feat(moe): MoE + MolT5 LLM 融合层、消融流水线

lnp_ml/dataset.py

@@ -139,6 +139,12 @@ def process_dataframe(df: pd.DataFrame) -> pd.DataFrame:
     for col in TARGET_BIODIST:
         if col in df.columns:
             df[col] = pd.to_numeric(df[col], errors="coerce").fillna(0.0)
+    if all(col in df.columns for col in TARGET_BIODIST):
+        bd = df[TARGET_BIODIST].values.astype(float)
+        s = bd.sum(axis=1, keepdims=True)
+        nz = s.squeeze(-1) > 0
+        bd[nz] = bd[nz] / s[nz]
+        df[TARGET_BIODIST] = bd
     
     return df
 

lnp_ml/modeling/heads.py

@@ -94,6 +94,12 @@ class MultiTaskHead(nn.Module):
         # toxic: 二分类
         self.toxic_head = ClassificationHead(in_dim, num_classes=2, hidden_dim=hidden_dim, dropout=dropout)
 
+        # 不确定性加权（Kendall 2018）：每个任务一个可学习 log σ²，初始 0
+        self.log_vars = nn.ParameterDict({
+            t: nn.Parameter(torch.zeros(()))
+            for t in ["size", "delivery", "pdi", "ee", "toxic", "biodist"]
+        })
+
     def forward(self, x: torch.Tensor) -> Dict[str, torch.Tensor]:
         """
         Args:

lnp_ml/modeling/layers/fusion.py

@@ -1,7 +1,7 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from typing import Dict, Literal, Tuple, Union
+from typing import Dict, List, Literal, Optional, Tuple, Union
 
 
 PoolingStrategy = Literal["concat", "avg", "max", "attention"]
@@ -109,3 +109,44 @@ class FusionLayer(nn.Module):
         if return_attn_weights:
             return out, attn_weights.squeeze(1)  # [B, n_tokens]
         return out
+
+
+class ResidualConcatFusion(nn.Module):
+    """对真实 token 做 attention pooling，再用零初始化门把 MoE/LLM 旁路以残差方式加入。
+
+    g_moe / g_llm 初始为 0 → +moe/+llm 起点严格等于 baseline；
+    旁路只有确实有用时才会被训练打开，从机制上保证“加了不会更差”。
+    """
+
+    def __init__(self, d_model: int, strategy: PoolingStrategy = "attention") -> None:
+        super().__init__()
+        if strategy == "concat":
+            raise ValueError("ResidualConcatFusion 不支持 concat（token 数随开关变化）")
+        self.d_model = d_model
+        self.pool = FusionLayer(d_model=d_model, n_tokens=1, strategy=strategy)
+        self.fusion_dim = self.pool.fusion_dim
+        # 零初始化门控（可学习标量），旁路初始不参与
+        self.g_moe = nn.Parameter(torch.zeros(()))
+        self.g_llm = nn.Parameter(torch.zeros(()))
+
+    def forward(
+        self,
+        chem: torch.Tensor,
+        tab: torch.Tensor,
+        f_moe: Optional[torch.Tensor] = None,
+        f_llm: Optional[torch.Tensor] = None,
+        return_attn_weights: bool = False,
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        # 只对真实 token（chem + tab）做注意力池化，旁路不参与 softmax 竞争
+        seq = torch.cat([chem, tab], dim=1)  # [B, n_chem + n_cond, d_model]
+        pooled = self.pool(seq, return_attn_weights=return_attn_weights)
+        if return_attn_weights:
+            pooled, attn = pooled
+
+        out = pooled
+        if f_moe is not None:
+            out = out + self.g_moe * f_moe  # 残差 + 零初始化门
+        if f_llm is not None:
+            out = out + self.g_llm * f_llm
+
+        return (out, attn) if return_attn_weights else out

lnp_ml/modeling/layers/llm_prompt.py

@@ -0,0 +1,104 @@
+"""LLM 分子特征分支：用 MolT5 直接编码 SMILES 文本。"""
+
+import os
+from typing import Dict, List, Optional
+
+import torch
+import torch.nn as nn
+
+
+# 权重默认路径，可用环境变量 MOLT5_PATH 覆盖
+DEFAULT_MOLT5_PATH = os.environ.get("MOLT5_PATH", "models/molt5-base")
+
+
+class LLMPromptEncoder(nn.Module):
+    """用 MolT5 编码 SMILES 文本，输出分子特征 F_llm [B, d_model]。
+
+    - 冻结 encoder 时：对每个 SMILES 缓存其句向量，避免重复前向，降方差、提速。
+    - use_lora=True 时：encoder 可训练（LoRA），不缓存。
+    """
+
+    def __init__(
+        self,
+        d_model: int,
+        n_chem_tokens: int = 4,
+        n_cond_tokens: int = 4,
+        model_name_or_path: str = DEFAULT_MOLT5_PATH,
+        freeze: bool = True,
+        use_lora: bool = False,
+        lora_r: int = 8,
+        lora_alpha: int = 16,
+        lora_dropout: float = 0.05,
+        max_length: int = 128,
+    ) -> None:
+        super().__init__()
+        from transformers import AutoTokenizer, T5EncoderModel
+
+        self.use_lora = use_lora
+        self.max_length = max_length
+        self.tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
+        self.encoder = T5EncoderModel.from_pretrained(model_name_or_path)
+        self.hidden_size = self.encoder.config.d_model
+
+        if use_lora:
+            self._apply_lora(lora_r, lora_alpha, lora_dropout)
+        elif freeze:
+            for p in self.encoder.parameters():
+                p.requires_grad = False
+        self._frozen = freeze and not use_lora
+
+        self.proj_down = nn.Sequential(
+            nn.Linear(self.hidden_size, d_model), nn.LayerNorm(d_model)
+        )
+        # 冻结特征缓存：smiles -> [H]（CPU）
+        self._cache: Dict[str, torch.Tensor] = {}
+
+    def _apply_lora(self, r: int, alpha: int, dropout: float) -> None:
+        from peft import LoraConfig, get_peft_model
+
+        for p in self.encoder.parameters():
+            p.requires_grad = False
+        cfg = LoraConfig(
+            r=r, lora_alpha=alpha, lora_dropout=dropout,
+            target_modules=["q", "k", "v", "o"], bias="none",
+        )
+        self.encoder = get_peft_model(self.encoder, cfg)
+
+    def _mean_pool(self, last_hidden: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
+        m = mask.unsqueeze(-1).float()                       # [B, L, 1]
+        return (last_hidden * m).sum(1) / m.sum(1).clamp(min=1e-6)
+
+    @torch.no_grad()
+    def _encode_frozen(self, smiles: List[str], device: torch.device) -> torch.Tensor:
+        missing = [s for s in smiles if s not in self._cache]
+        if missing:
+            uniq = list(dict.fromkeys(missing))
+            for i in range(0, len(uniq), 256):
+                chunk = uniq[i:i + 256]
+                enc = self.tokenizer(
+                    chunk, padding=True, truncation=True,
+                    max_length=self.max_length, return_tensors="pt",
+                ).to(device)
+                out = self.encoder(**enc).last_hidden_state
+                pooled = self._mean_pool(out, enc["attention_mask"])
+                for s, v in zip(chunk, pooled):
+                    self._cache[s] = v.cpu()
+        return torch.stack([self._cache[s] for s in smiles]).to(device)
+
+    def _encode_trainable(self, smiles: List[str], device: torch.device) -> torch.Tensor:
+        enc = self.tokenizer(
+            list(smiles), padding=True, truncation=True,
+            max_length=self.max_length, return_tensors="pt",
+        ).to(device)
+        out = self.encoder(**enc).last_hidden_state
+        return self._mean_pool(out, enc["attention_mask"])
+
+    def forward(self, smiles: List[str], tab: Optional[torch.Tensor] = None) -> torch.Tensor:
+        """Args: smiles [B] SMILES 字符串列表。Returns: [B, d_model]。"""
+        device = self.proj_down[0].weight.device
+        feat = self._encode_frozen(smiles, device) if self._frozen \
+            else self._encode_trainable(smiles, device)
+        return self.proj_down(feat)
+
+    def clear_cache(self) -> None:
+        self._cache.clear()

lnp_ml/modeling/layers/set_transformer.py

@@ -0,0 +1,123 @@
+"""Set Transformer 集合编码器。
+输入/输出形状均为 [B, n_tokens, d_model]，
+n_tokens 可变（支持 3 或 4）。
+"""
+
+import torch
+import torch.nn as nn
+
+
+class MAB(nn.Module):
+    """多头注意力块：MAB(Q, K) = LN(H + FFN(H))，H = LN(Q + MultiHeadAttn(Q, K, K))。"""
+
+    def __init__(self, d_model: int, num_heads: int, dropout: float = 0.1, ln: bool = True) -> None:
+        super().__init__()
+        assert d_model % num_heads == 0, "d_model 必须能被 num_heads 整除"
+        self.attn = nn.MultiheadAttention(
+            d_model, num_heads, dropout=dropout, batch_first=True
+        )
+        self.norm1 = nn.LayerNorm(d_model) if ln else nn.Identity()
+        self.norm2 = nn.LayerNorm(d_model) if ln else nn.Identity()
+        self.ffn = nn.Sequential(
+            nn.Linear(d_model, d_model * 4),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(d_model * 4, d_model),
+            nn.Dropout(dropout),
+        )
+
+    def forward(self, q: torch.Tensor, k: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            q: [B, n_q, d_model]
+            k: [B, n_k, d_model]
+
+        Returns:
+            [B, n_q, d_model]
+        """
+        attn_out, _ = self.attn(q, k, k)
+        h = self.norm1(q + attn_out)
+        return self.norm2(h + self.ffn(h))
+
+
+class SAB(nn.Module):
+    """集合自注意力块：SAB(X) = MAB(X, X)。"""
+
+    def __init__(self, d_model: int, num_heads: int, dropout: float = 0.1, ln: bool = True) -> None:
+        super().__init__()
+        self.mab = MAB(d_model, num_heads, dropout, ln)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """[B, n, d_model] -> [B, n, d_model]"""
+        return self.mab(x, x)
+
+
+class ISAB(nn.Module):
+    """诱导点集合注意力块：用 m 个可学习诱导点降低大集合的注意力复杂度。"""
+
+    def __init__(
+        self,
+        d_model: int,
+        num_heads: int,
+        num_inducing: int = 16,
+        dropout: float = 0.1,
+        ln: bool = True,
+    ) -> None:
+        super().__init__()
+        # 可学习诱导点
+        self.inducing = nn.Parameter(torch.empty(1, num_inducing, d_model))
+        nn.init.xavier_uniform_(self.inducing)
+        self.mab_in = MAB(d_model, num_heads, dropout, ln)
+        self.mab_out = MAB(d_model, num_heads, dropout, ln)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """[B, n, d_model] -> [B, n, d_model]"""
+        inducing = self.inducing.expand(x.size(0), -1, -1)  # [B, m, d_model]
+        h = self.mab_in(inducing, x)  # [B, m, d_model]
+        return self.mab_out(x, h)     # [B, n, d_model]
+
+
+class SetTransformer(nn.Module):
+    """对化学 token 集合做 N 层集合自注意力，形状保持 [B, n_tokens, d_model]。"""
+
+    def __init__(
+        self,
+        d_model: int,
+        num_heads: int = 8,
+        n_layers: int = 4,
+        dropout: float = 0.1,
+        block: str = "sab",
+        num_inducing: int = 16,
+        ln: bool = True,
+    ) -> None:
+        """
+        Args:
+            d_model: token 维度
+            num_heads: 注意力头数，d_head = d_model / num_heads
+            n_layers: 集合注意力层数
+            dropout: dropout 比例
+            block: 注意力块类型，"sab"（全自注意力）或 "isab"（诱导点）
+            num_inducing: ISAB 的诱导点数量（block="isab" 时生效）
+            ln: 是否使用 LayerNorm
+        """
+        super().__init__()
+        self.block_type = block
+
+        self.layers = nn.ModuleList()
+        for _ in range(n_layers):
+            if block == "isab":
+                self.layers.append(ISAB(d_model, num_heads, num_inducing, dropout, ln))
+            else:
+                self.layers.append(SAB(d_model, num_heads, dropout, ln))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            x: [B, n_tokens, d_model] 化学 token 集合
+
+        Returns:
+            [B, n_tokens, d_model] 集合编码后的 token
+        """
+        for layer in self.layers:
+            x = layer(x)
+        return x

lnp_ml/modeling/models.py

@@ -5,14 +5,21 @@ import torch.nn as nn
 from typing import Dict, List, Optional, Literal
 
 from lnp_ml.modeling.encoders import CachedRDKitEncoder, CachedMPNNEncoder
-from lnp_ml.modeling.layers import TokenProjector, CrossModalAttention, FusionLayer, MoEBlock
+from lnp_ml.modeling.layers import (
+    TokenProjector,
+    SetTransformer,
+    ResidualConcatFusion,
+    MoEBlock,
+    LLMPromptEncoder,
+)
+from lnp_ml.modeling.layers.llm_prompt import DEFAULT_MOLT5_PATH
 from lnp_ml.modeling.heads import MultiTaskHead
 
 
-PoolingStrategy = Literal["concat", "avg", "max", "attention"]
+PoolingStrategy = Literal["attention", "avg", "max"]
 
 
-# Token 维度配置（根据 ARCHITECTURE.md）
+# Token 维度配置
 DEFAULT_INPUT_DIMS = {
     # Channel A: 化学特征
     "mpnn": 600,      # D-MPNN embedding
@@ -26,8 +33,21 @@ DEFAULT_INPUT_DIMS = {
     "exp": 32,        # 实验条件 one-hot
 }
 
-# Token 顺序（前 4 个为 Channel A，后 4 个为 Channel B）
-TOKEN_ORDER = ["mpnn", "morgan", "maccs", "desc", "comp", "phys", "help", "exp"]
+# 化学 / 配方 token 的键顺序
+CHEM_KEYS_WITH_MPNN = ["mpnn", "morgan", "maccs", "desc"]
+CHEM_KEYS_NO_MPNN = ["morgan", "maccs", "desc"]
+TAB_KEYS = ["comp", "phys", "help", "exp"]
+
+# backbone 权重前缀（用于预训练加载与导出）
+BACKBONE_PREFIXES = (
+    "token_projector.",
+    "set_transformer.",
+    "fusion.",
+    "moe.",
+    "llm_prompt.",
+)
+# 冻结的 MolT5 encoder 权重前缀，不纳入 backbone（由本地权重加载，不进 checkpoint）
+LLM_FROZEN_PREFIX = "llm_prompt.encoder."
 
 
 class LNPModel(nn.Module):
@@ -37,37 +57,47 @@ class LNPModel(nn.Module):
     架构流程:
         1. Encoders: SMILES -> 化学特征; tabular -> 配方/实验特征
         2. TokenProjector: 统一到 d_model
-        3. Stack: [B, 8, d_model]
-        4. CrossModalAttention: Channel A (化学) <-> Channel B (配方/实验)
-        5. FusionLayer: [B, 8, d_model] -> [B, fusion_dim]
-        6. MultiTaskHead: 多任务预测
+        3. SetTransformer: 对化学 token 集合做置换等变编码 -> chem'
+        4. MoE (可选): router 看 tab，expert 吃 chem' -> F_moe
+        5. LLM (可选): chem'/tab 注入 MolT5 prompt -> F_llm
+        6. ResidualConcatFusion: 拼接 chem'/tab/F_moe/F_llm -> attention pooling
+        7. MultiTaskHead: 多任务预测
     """
 
     def __init__(
         self,
         # 模型维度
         d_model: int = 256,
-        # Cross attention
+        # Set Transformer
         num_heads: int = 8,
         n_attn_layers: int = 4,
+        set_transformer_block: str = "sab",
         # Fusion
         fusion_strategy: PoolingStrategy = "attention",
         # Head
         head_hidden_dim: int = 128,
         # Dropout
         dropout: float = 0.1,
-        # MPNN encoder (可选，如果不用 MPNN 可以设为 None)
+        # MPNN encoder
         mpnn_checkpoint: Optional[str] = None,
         mpnn_ensemble_paths: Optional[List[str]] = None,
         mpnn_device: str = "cpu",
         # 输入维度配置
         input_dims: Optional[Dict[str, int]] = None,
-        # ============ MoE 相关（新增） ============
+        # ============ MoE 相关 ============
         use_moe: bool = False,
         moe_n_experts: int = 4,
         moe_top_k: int = 2,
         moe_expert_hidden_mult: int = 2,
         moe_jitter_noise: float = 0.0,
+        # ============ LLM 相关 ============
+        use_llm: bool = False,
+        llm_model_path: str = DEFAULT_MOLT5_PATH,
+        llm_freeze: bool = True,
+        llm_use_lora: bool = False,
+        llm_lora_r: int = 8,
+        llm_lora_alpha: int = 16,
+        llm_lora_dropout: float = 0.05,
     ) -> None:
         super().__init__()
 
@@ -76,10 +106,7 @@ class LNPModel(nn.Module):
         self.use_mpnn = mpnn_checkpoint is not None or mpnn_ensemble_paths is not None
 
         # ============ Encoders ============
-        # RDKit encoder (always used)
         self.rdkit_encoder = CachedRDKitEncoder()
-
-        # MPNN encoder (optional)
         if self.use_mpnn:
             self.mpnn_encoder = CachedMPNNEncoder(
                 checkpoint_path=mpnn_checkpoint,
@@ -90,27 +117,28 @@ class LNPModel(nn.Module):
             self.mpnn_encoder = None
 
         # ============ Token Projector ============
-        # 根据是否使用 MPNN 调整输入维度
         proj_input_dims = {k: v for k, v in self.input_dims.items()}
         if not self.use_mpnn:
             proj_input_dims.pop("mpnn", None)
-
         self.token_projector = TokenProjector(
             input_dims=proj_input_dims,
             d_model=d_model,
             dropout=dropout,
         )
 
-        # ============ Cross Modal Attention ============
-        n_tokens = 8 if self.use_mpnn else 7
-        split_idx = 4 if self.use_mpnn else 3  # Channel A 的 token 数量
+        # token 顺序与化学侧 token 数
+        self.chem_keys = CHEM_KEYS_WITH_MPNN if self.use_mpnn else CHEM_KEYS_NO_MPNN
+        self.tab_keys = TAB_KEYS
+        self.token_order = self.chem_keys + self.tab_keys
+        self.split_idx = len(self.chem_keys)
 
-        self.cross_attention = CrossModalAttention(
+        # ============ Set Transformer ============
+        self.set_transformer = SetTransformer(
             d_model=d_model,
             num_heads=num_heads,
             n_layers=n_attn_layers,
-            split_idx=split_idx,
             dropout=dropout,
+            block=set_transformer_block,
         )
 
         # ============ MoE Block (可选) ============
@@ -119,24 +147,35 @@ class LNPModel(nn.Module):
         if use_moe:
             self.moe = MoEBlock(
                 d_model=d_model,
-                n_chem_tokens=split_idx,
+                n_chem_tokens=self.split_idx,
                 n_experts=moe_n_experts,
                 top_k=moe_top_k,
                 expert_hidden_mult=moe_expert_hidden_mult,
                 dropout=dropout,
                 jitter_noise=moe_jitter_noise,
             )
-            n_fusion_tokens = n_tokens + 1  # 多一个 F_moe token
         else:
             self.moe = None
-            n_fusion_tokens = n_tokens
 
-        # ============ Fusion Layer ============
-        self.fusion = FusionLayer(
-            d_model=d_model,
-            n_tokens=n_fusion_tokens,
-            strategy=fusion_strategy,
-        )
+        # ============ LLM Prompt (可选) ============
+        self.use_llm = use_llm
+        if use_llm:
+            self.llm_prompt = LLMPromptEncoder(
+                d_model=d_model,
+                n_chem_tokens=self.split_idx,
+                n_cond_tokens=len(self.tab_keys),
+                model_name_or_path=llm_model_path,
+                freeze=llm_freeze,
+                use_lora=llm_use_lora,
+                lora_r=llm_lora_r,
+                lora_alpha=llm_lora_alpha,
+                lora_dropout=llm_lora_dropout,
+            )
+        else:
+            self.llm_prompt = None
+
+        # ============ Residual Concat + Fusion ============
+        self.fusion = ResidualConcatFusion(d_model=d_model, strategy=fusion_strategy)
 
         # ============ Multi-Task Head ============
         self.head = MultiTaskHead(
@@ -151,70 +190,51 @@ class LNPModel(nn.Module):
         tabular: Dict[str, torch.Tensor],
     ) -> torch.Tensor:
         """
-        内部方法：编码 SMILES 和 tabular，返回 stacked tokens。
+        编码 SMILES 和 tabular，返回 stacked tokens。
 
         Returns:
-            stacked: [B, n_tokens, d_model]
+            stacked: [B, n_tokens, d_model]，顺序为 chem 在前、tab 在后
         """
-        # 获取目标设备（从 tabular 数据推断）
         device = tabular["comp"].device
 
-        # 1. Encode SMILES
-        rdkit_features = self.rdkit_encoder(smiles)  # {"morgan", "maccs", "desc"}
+        rdkit_features = self.rdkit_encoder(smiles)
 
-        # 2. 合并所有特征
         all_features: Dict[str, torch.Tensor] = {}
-
-        # MPNN 特征（如果启用）
         if self.use_mpnn:
             mpnn_features = self.mpnn_encoder(smiles)
             all_features["mpnn"] = mpnn_features["mpnn"].to(device)
-
-        # RDKit 特征（移到正确设备）
         all_features["morgan"] = rdkit_features["morgan"].to(device)
         all_features["maccs"] = rdkit_features["maccs"].to(device)
         all_features["desc"] = rdkit_features["desc"].to(device)
-
-        # Tabular 特征（已在正确设备上）
         all_features["comp"] = tabular["comp"]
         all_features["phys"] = tabular["phys"]
         all_features["help"] = tabular["help"]
         all_features["exp"] = tabular["exp"]
 
-        # 3. Token Projector: 统一维度
-        projected = self.token_projector(all_features)  # Dict[str, [B, d_model]]
-
-        # 4. Stack tokens: [B, n_tokens, d_model]
-        if self.use_mpnn:
-            token_order = ["mpnn", "morgan", "maccs", "desc", "comp", "phys", "help", "exp"]
-        else:
-            token_order = ["morgan", "maccs", "desc", "comp", "phys", "help", "exp"]
-
-        stacked = torch.stack([projected[k] for k in token_order], dim=1)
+        projected = self.token_projector(all_features)
+        stacked = torch.stack([projected[k] for k in self.token_order], dim=1)
         return stacked
 
-    def _attended_with_moe(self, stacked: torch.Tensor) -> torch.Tensor:
-        """
-        Cross-attention + 可选 MoE 旁路 → fusion 输入序列。
+    def _backbone_from_stacked(
+        self, stacked: torch.Tensor, smiles: Optional[List[str]] = None
+    ) -> torch.Tensor:
+        chem = stacked[:, : self.split_idx, :]
+        tab = stacked[:, self.split_idx :, :]
 
-        - 不启用 MoE 时：返回 [B, n_tokens, d]，与原行为一致。
-        - 启用 MoE 时：在最后追加一个 F_moe token，返回 [B, n_tokens + 1, d]。
+        chem = self.set_transformer(chem)  # chem'
 
-        副作用:
-            把本次 forward 的 MoE 副产物（aux loss / gates / probs）写到
-            self._last_moe_extras，trainer 端通过 get_last_moe_extras() 读取。
-        """
-        attended = self.cross_attention(stacked)
+        f_moe = None
         if self.moe is not None:
-            split = self.cross_attention.split_idx
-            chem_prime = attended[:, :split, :]
-            tab_prime = attended[:, split:, :]
-            F_moe, extras = self.moe(chem_prime, tab_prime)
+            f_moe, extras = self.moe(chem, tab)
             self._last_moe_extras = extras
-            attended = torch.cat([attended, F_moe.unsqueeze(1)], dim=1)
         else:
             self._last_moe_extras = None
-        return attended
+
+        f_llm = None
+        if self.llm_prompt is not None and smiles is not None:
+            f_llm = self.llm_prompt(smiles)
+
+        return self.fusion(chem, tab, f_moe=f_moe, f_llm=f_llm)
 
     def forward_from_projected(
         self,
@@ -225,15 +245,13 @@ class LNPModel(nn.Module):
         从已投影的 stacked tokens 开始 forward，用于 Captum 归因。
 
         Args:
-            stacked: [B, n_tokens, d_model] TokenProjector 输出后 stack 的张量。
-            task: 指定单任务名 ("size", "pdi", "ee", "delivery", "biodist", "toxic")。
-                  若为 None，返回 delivery head 的标量输出。
+            stacked: [B, n_tokens, d_model]
+            task: 单任务名；None 时返回 delivery head 输出。
 
         Returns:
-            [B, 1] 或 [B, num_classes] 对应任务的预测输出。
+            对应任务的预测输出。
         """
-        attended = self._attended_with_moe(stacked)
-        fused = self.fusion(attended)
+        fused = self._backbone_from_stacked(stacked)
 
         if task is None:
             task = "delivery"
@@ -259,19 +277,10 @@ class LNPModel(nn.Module):
         用原始特征替换 base_projected 中指定 token 的投影，然后 forward。
 
         用于对单个 token 内部特征做 Captum 归因（如 desc 的 210 维）。
-
-        Args:
-            raw_feature: [B, input_dim] 某个 token 的原始特征
-            feature_key: token 名称，如 "desc"
-            base_projected: [B, n_tokens, d_model] 其他 token 已投影好的张量
-            task: 任务名
-
-        Returns:
-            对应任务的预测输出
         """
         projected = self.token_projector.projectors[feature_key](raw_feature)
         gate = torch.sigmoid(self.token_projector.weights[feature_key])
-        projected = projected * gate  # [B, d_model]
+        projected = projected * gate
 
         token_order = list(self.token_projector.keys)
         token_idx = token_order.index(feature_key)
@@ -286,38 +295,16 @@ class LNPModel(nn.Module):
         smiles: List[str],
         tabular: Dict[str, torch.Tensor],
     ) -> torch.Tensor:
-        """
-        Backbone forward：编码 -> 投影 -> 注意力 -> (可选 MoE) -> 融合，不经过任务头。
-
-        用于 pretrain 阶段或需要提取特征的场景。
-
-        Args:
-            smiles: SMILES 字符串列表，长度为 B
-            tabular: Dict[str, Tensor]
-
-        Returns:
-            fused: [B, fusion_dim] 融合后的特征向量
-        """
+        """Backbone forward：编码 -> 投影 -> set transformer -> (MoE/LLM) -> 融合。"""
         stacked = self._encode_and_project(smiles, tabular)
-        attended = self._attended_with_moe(stacked)
-        fused = self.fusion(attended)
-        return fused
+        return self._backbone_from_stacked(stacked, smiles=smiles)
 
     def forward_delivery(
         self,
         smiles: List[str],
         tabular: Dict[str, torch.Tensor],
     ) -> torch.Tensor:
-        """
-        仅预测 delivery（用于 pretrain）。
-        
-        Args:
-            smiles: SMILES 字符串列表，长度为 B
-            tabular: Dict[str, Tensor]
-            
-        Returns:
-            delivery: [B, 1] 预测的 delivery 值
-        """
+        """仅预测 delivery（用于 pretrain）。返回 [B, 1]。"""
         fused = self.forward_backbone(smiles, tabular)
         return self.head.delivery_head(fused)
 
@@ -329,52 +316,33 @@ class LNPModel(nn.Module):
         """
         完整的多任务 forward。
 
-        Args:
-            smiles: SMILES 字符串列表，长度为 B
-            tabular: Dict[str, Tensor]，包含:
-                - "comp": [B, 5] 配方比例
-                - "phys": [B, 12] 物理参数
-                - "help": [B, 4] Helper lipid
-                - "exp": [B, 32] 实验条件
-
         Returns:
-            Dict[str, Tensor]:
-                - "size": [B, 1]
-                - "pdi": [B, 4]
-                - "ee": [B, 3]
-                - "delivery": [B, 1]
-                - "biodist": [B, 7]
-                - "toxic": [B, 2]
+            Dict[str, Tensor]: size [B,1], pdi [B,4], ee [B,3],
+            delivery [B,1], biodist [B,7], toxic [B,2]
         """
         fused = self.forward_backbone(smiles, tabular)
-        outputs = self.head(fused)
-        return outputs
+        return self.head(fused)
 
     def clear_cache(self) -> None:
         """清空所有 encoder 的缓存"""
         self.rdkit_encoder.clear_cache()
         if self.mpnn_encoder is not None:
             self.mpnn_encoder.clear_cache()
+        if self.llm_prompt is not None and hasattr(self.llm_prompt, "clear_cache"):
+            self.llm_prompt.clear_cache()
 
     def get_last_moe_extras(self) -> Optional[Dict[str, torch.Tensor]]:
-        """返回最近一次 forward 中 MoE 模块的副产物（aux loss、gates、probs）。
-
-        若未启用 MoE 或还未调用过 forward，返回 None。
-        """
+        """返回最近一次 forward 中 MoE 模块的副产物（aux loss、gates、probs）。"""
         return self._last_moe_extras
 
     def get_backbone_state_dict(self) -> Dict[str, torch.Tensor]:
         """
-        获取 backbone 部分的 state_dict（不含任务头）。
-
-        包含: token_projector, cross_attention, fusion，以及（启用时）moe。
+        获取 backbone 部分的 state_dict（不含任务头，且排除冻结的 MolT5 encoder）。
         """
-        backbone_prefixes = ("token_projector.", "cross_attention.", "fusion.", "moe.")
-        backbone_keys = [
-            name for name in self.state_dict().keys()
-            if name.startswith(backbone_prefixes)
-        ]
-        return {k: v for k, v in self.state_dict().items() if k in backbone_keys}
+        return {
+            k: v for k, v in self.state_dict().items()
+            if k.startswith(BACKBONE_PREFIXES) and not k.startswith(LLM_FROZEN_PREFIX)
+        }
 
     def get_delivery_head_state_dict(self) -> Dict[str, torch.Tensor]:
         """获取 delivery head 的 state_dict"""
@@ -392,16 +360,11 @@ class LNPModel(nn.Module):
         """
         从预训练 checkpoint 加载 backbone 和（可选）delivery head 权重。
 
-        Args:
-            pretrain_state_dict: 预训练模型的 state_dict
-            load_delivery_head: 是否加载 delivery head 权重
-            strict: 是否严格匹配（默认 False，允许缺失/多余的键）
+        冻结的 MolT5 encoder 权重不在加载范围内。
         """
-        backbone_prefixes = ("token_projector.", "cross_attention.", "fusion.", "moe.")
-
         keys_to_load = []
         for name in pretrain_state_dict.keys():
-            if name.startswith(backbone_prefixes):
+            if name.startswith(BACKBONE_PREFIXES) and not name.startswith(LLM_FROZEN_PREFIX):
                 keys_to_load.append(name)
             elif load_delivery_head and name.startswith("head.delivery_head."):
                 keys_to_load.append(name)
@@ -410,45 +373,49 @@ class LNPModel(nn.Module):
             k: v for k, v in pretrain_state_dict.items() if k in keys_to_load
         }
 
-        missing, unexpected = [], []
+        unexpected = []
         model_state = self.state_dict()
         for k, v in filtered_state_dict.items():
-            if k in model_state:
-                if model_state[k].shape == v.shape:
-                    model_state[k] = v
-                else:
-                    unexpected.append(
-                        f"{k} (shape mismatch: {model_state[k].shape} vs {v.shape})"
-                    )
+            if k in model_state and model_state[k].shape == v.shape:
+                model_state[k] = v
             else:
                 unexpected.append(k)
 
+
         self.load_state_dict(model_state, strict=False)
 
-        if strict and (missing or unexpected):
-            raise RuntimeError(f"Missing keys: {missing}, Unexpected keys: {unexpected}")
+        if strict and unexpected:
+            raise RuntimeError(f"Unexpected keys: {unexpected}")
 
 
 class LNPModelWithoutMPNN(LNPModel):
-    """不使用 MPNN 的简化版本"""
+    """不使用 MPNN 的简化版本（化学 token 为 3 个）"""
 
     def __init__(
         self,
         d_model: int = 256,
         num_heads: int = 8,
         n_attn_layers: int = 4,
+        set_transformer_block: str = "sab",
         fusion_strategy: PoolingStrategy = "attention",
         head_hidden_dim: int = 128,
         dropout: float = 0.1,
         input_dims: Optional[Dict[str, int]] = None,
-        # ============ MoE 相关（新增） ============
+        # ============ MoE 相关 ============
         use_moe: bool = False,
         moe_n_experts: int = 4,
         moe_top_k: int = 2,
         moe_expert_hidden_mult: int = 2,
         moe_jitter_noise: float = 0.0,
+        # ============ LLM 相关 ============
+        use_llm: bool = False,
+        llm_model_path: str = DEFAULT_MOLT5_PATH,
+        llm_freeze: bool = True,
+        llm_use_lora: bool = False,
+        llm_lora_r: int = 8,
+        llm_lora_alpha: int = 16,
+        llm_lora_dropout: float = 0.05,
     ) -> None:
-        # 移除 mpnn 维度
         dims = input_dims or DEFAULT_INPUT_DIMS.copy()
         dims.pop("mpnn", None)
 
@@ -456,6 +423,7 @@ class LNPModelWithoutMPNN(LNPModel):
             d_model=d_model,
             num_heads=num_heads,
             n_attn_layers=n_attn_layers,
+            set_transformer_block=set_transformer_block,
             fusion_strategy=fusion_strategy,
             head_hidden_dim=head_hidden_dim,
             dropout=dropout,
@@ -467,5 +435,11 @@ class LNPModelWithoutMPNN(LNPModel):
             moe_top_k=moe_top_k,
             moe_expert_hidden_mult=moe_expert_hidden_mult,
             moe_jitter_noise=moe_jitter_noise,
+            use_llm=use_llm,
+            llm_model_path=llm_model_path,
+            llm_freeze=llm_freeze,
+            llm_use_lora=llm_use_lora,
+            llm_lora_r=llm_lora_r,
+            llm_lora_alpha=llm_lora_alpha,
+            llm_lora_dropout=llm_lora_dropout,
         )
-

lnp_ml/modeling/nested_cv_optuna.py

@@ -44,6 +44,8 @@ from lnp_ml.dataset import (
 from tqdm import tqdm
 
 from lnp_ml.modeling.models import LNPModel, LNPModelWithoutMPNN
+from lnp_ml.modeling.layers.llm_prompt import DEFAULT_MOLT5_PATH
+from lnp_ml.utils.seed import set_global_seed
 from lnp_ml.modeling.encoders.rdkit_encoder import CachedRDKitEncoder
 from lnp_ml.modeling.trainer_balanced import (
     ClassWeights,
@@ -178,19 +180,25 @@ def create_model(
     dropout: float = 0.1,
     use_mpnn: bool = False,
     mpnn_device: str = "cpu",
+    set_transformer_block: str = "sab",
+    # MoE
     use_moe: bool = False,
     moe_n_experts: int = 4,
     moe_top_k: int = 2,
     moe_expert_hidden_mult: int = 2,
     moe_jitter_noise: float = 0.0,
+    # LLM（打包成 dict 以减少多进程参数传递）
+    llm_kwargs: Optional[Dict] = None,
 ) -> Union[LNPModel, LNPModelWithoutMPNN]:
-    """创建模型"""
-    moe_kwargs = dict(
+    """创建模型。llm_kwargs 含 use_llm / llm_model_path / llm_freeze / llm_use_lora / llm_lora_*。"""
+    extra_kwargs = dict(
+        set_transformer_block=set_transformer_block,
         use_moe=use_moe,
         moe_n_experts=moe_n_experts,
         moe_top_k=moe_top_k,
         moe_expert_hidden_mult=moe_expert_hidden_mult,
         moe_jitter_noise=moe_jitter_noise,
+        **(llm_kwargs or {}),
     )
 
     if use_mpnn:
@@ -204,7 +212,7 @@ def create_model(
             dropout=dropout,
             mpnn_ensemble_paths=ensemble_paths,
             mpnn_device=mpnn_device,
-            **moe_kwargs,
+            **extra_kwargs,
         )
     else:
         return LNPModelWithoutMPNN(
@@ -214,7 +222,7 @@ def create_model(
             fusion_strategy=fusion_strategy,
             head_hidden_dim=head_hidden_dim,
             dropout=dropout,
-            **moe_kwargs,
+            **extra_kwargs,
         )
 
 
@@ -381,6 +389,8 @@ def run_inner_optuna(
     moe_top_k: int = 2,
     moe_expert_hidden_mult: int = 2,
     moe_jitter_noise: float = 0.0,
+    set_transformer_block: str = "sab",
+    llm_kwargs: Optional[Dict] = None,
 ) -> Tuple[Dict, int, optuna.Study]:
     """
     在内层数据上运行 Optuna 超参搜索。
@@ -436,6 +446,16 @@ def run_inner_optuna(
         weight_decay = trial.suggest_float("weight_decay", 1e-5, 1e-1, log=True)
         backbone_lr_ratio = trial.suggest_float("backbone_lr_ratio", 0.01, 1.0, log=True)
         
+        # MoE / LLM
+        use_llm = bool(llm_kwargs.get("use_llm", False))
+        moe_ne_t = trial.suggest_categorical("moe_n_experts", [2, 4]) if use_moe else moe_n_experts
+        moe_tk_t = trial.suggest_int("moe_top_k", 1, 2) if use_moe else moe_top_k
+        moe_hm_t = trial.suggest_categorical("moe_expert_hidden_mult", [1, 2]) if use_moe else moe_expert_hidden_mult
+        llm_kwargs_t = dict(llm_kwargs)
+        if use_llm:
+            llm_kwargs_t["llm_use_lora"] = True
+            llm_kwargs_t["llm_lora_r"] = trial.suggest_categorical("llm_lora_r", [8, 16, 32])
+
         # 内层 3-fold CV
         inner_cv = StratifiedKFold(
             n_splits=n_inner_folds, shuffle=True, random_state=seed
@@ -476,10 +496,12 @@ def run_inner_optuna(
                 use_mpnn=use_mpnn,
                 mpnn_device=device.type,
                 use_moe=use_moe,
-                moe_n_experts=moe_n_experts,
-                moe_top_k=moe_top_k,
-                moe_expert_hidden_mult=moe_expert_hidden_mult,
+                moe_n_experts=moe_ne_t,
+                moe_top_k=moe_tk_t,
+                moe_expert_hidden_mult=moe_hm_t,
                 moe_jitter_noise=moe_jitter_noise,
+                set_transformer_block=set_transformer_block,
+                llm_kwargs=llm_kwargs_t,
             )
             if rdkit_cache is not None:
                 model.rdkit_encoder._cache = rdkit_cache
@@ -536,6 +558,7 @@ def run_inner_optuna(
         "n_attn_layers": fixed_n_attn_layers,
         "fusion_strategy": fixed_fusion_strategy,
         "head_hidden_dim": fixed_head_hidden_dim,
+        "set_transformer_block": set_transformer_block,
     })
     epoch_mean = study.best_trial.user_attrs.get("epoch_mean", epochs_per_trial)
     
@@ -571,6 +594,10 @@ def _run_single_outer_fold(
     moe_top_k: int = 2,
     moe_expert_hidden_mult: int = 2,
     moe_jitter_noise: float = 0.0,
+    set_transformer_block: str = "sab",
+    llm_kwargs: Optional[Dict] = None,
+    precomputed_best_params: Optional[Dict] = None,
+    precomputed_epoch_mean: Optional[int] = None,
 ) -> Dict:
     """
     执行单个外层 fold 的完整流程（内层调参 + 外层训练 + 评估）。
@@ -578,6 +605,7 @@ def _run_single_outer_fold(
     所有参数均为可序列化类型，以支持 spawn 多进程。
     """
     device = torch.device(device_str)
+    set_global_seed(seed + outer_fold)
     fold_dir = Path(fold_dir)
     fold_dir.mkdir(parents=True, exist_ok=True)
 
@@ -598,50 +626,66 @@ def _run_single_outer_fold(
             "outer_test_idx": outer_test_idx.tolist(),
         }, f)
 
-    # 内层 Optuna 调参
-    logger.info(f"\nRunning inner Optuna with {n_trials} trials...")
-    study_path = fold_dir / "optuna_study.sqlite3"
-
-    best_params, epoch_mean, study = run_inner_optuna(
-        full_dataset=full_dataset,
-        inner_train_indices=outer_train_idx,
-        strata=strata,
-        device=device,
-        n_trials=n_trials,
-        epochs_per_trial=epochs_per_trial,
-        patience=inner_patience,
-        batch_size=batch_size,
-        n_inner_folds=n_inner_folds,
-        use_mpnn=use_mpnn,
-        seed=seed + outer_fold,
-        study_path=study_path,
-        pretrain_state_dict=pretrain_state_dict,
-        pretrain_config=pretrain_config,
-        load_delivery_head=load_delivery_head,
-        rdkit_cache=rdkit_cache,
-        use_moe=use_moe,
-        moe_n_experts=moe_n_experts,
-        moe_top_k=moe_top_k,
-        moe_expert_hidden_mult=moe_expert_hidden_mult,
-        moe_jitter_noise=moe_jitter_noise,
-    )
+    # 内层 Optuna 调参（方式 B：已有超参则跳过，直接复用）
+    if precomputed_best_params is not None and precomputed_epoch_mean is not None:
+        logger.info("Reusing precomputed best_params (skip inner Optuna).")
+        best_params = dict(precomputed_best_params)
+        epoch_mean = int(precomputed_epoch_mean)
+    else:
+        logger.info(f"\nRunning inner Optuna with {n_trials} trials...")
+        study_path = fold_dir / "optuna_study.sqlite3"
+        best_params, epoch_mean, study = run_inner_optuna(
+            full_dataset=full_dataset,
+            inner_train_indices=outer_train_idx,
+            strata=strata,
+            device=device,
+            n_trials=n_trials,
+            epochs_per_trial=epochs_per_trial,
+            patience=inner_patience,
+            batch_size=batch_size,
+            n_inner_folds=n_inner_folds,
+            use_mpnn=use_mpnn,
+            seed=seed + outer_fold,
+            study_path=study_path,
+            pretrain_state_dict=pretrain_state_dict,
+            pretrain_config=pretrain_config,
+            load_delivery_head=load_delivery_head,
+            rdkit_cache=rdkit_cache,
+            use_moe=use_moe,
+            moe_n_experts=moe_n_experts,
+            moe_top_k=moe_top_k,
+            moe_expert_hidden_mult=moe_expert_hidden_mult,
+            moe_jitter_noise=moe_jitter_noise,
+            set_transformer_block=set_transformer_block,
+            llm_kwargs=llm_kwargs,
+        )
 
     # 保存最佳参数
     with open(fold_dir / "best_params.json", "w") as f:
         json.dump(best_params, f, indent=2)
-
     with open(fold_dir / "epoch_mean.json", "w") as f:
         json.dump({"epoch_mean": epoch_mean}, f)
 
     # 外层训练（使用最优超参，固定 epoch 数，不 early-stop）
     logger.info(f"\nTraining outer fold with best params, epochs={epoch_mean}...")
 
-    train_subset = Subset(full_dataset, outer_train_idx.tolist())
+    # 从 outer_train 再切 10% 作为"监控用 val"（只画曲线、看 gap，不参与选模型/早停）
+    _rng = np.random.RandomState(seed + outer_fold)
+    _perm = _rng.permutation(len(outer_train_idx))
+    _n_val = max(1, int(0.1 * len(outer_train_idx)))
+    fit_idx = outer_train_idx[_perm[_n_val:]]
+    monitor_idx = outer_train_idx[_perm[:_n_val]]
+
+    train_subset = Subset(full_dataset, fit_idx.tolist())
+    monitor_subset = Subset(full_dataset, monitor_idx.tolist())
     test_subset = Subset(full_dataset, outer_test_idx.tolist())
 
     train_loader = DataLoader(
         train_subset, batch_size=batch_size, shuffle=True, collate_fn=collate_fn
     )
+    monitor_loader = DataLoader(
+        monitor_subset, batch_size=batch_size, shuffle=False, collate_fn=collate_fn
+    )
     test_loader = DataLoader(
         test_subset, batch_size=batch_size, shuffle=False, collate_fn=collate_fn
     )
@@ -658,10 +702,13 @@ def _run_single_outer_fold(
         use_mpnn=use_mpnn,
         mpnn_device=device.type,
         use_moe=use_moe,
-        moe_n_experts=moe_n_experts,
-        moe_top_k=moe_top_k,
-        moe_expert_hidden_mult=moe_expert_hidden_mult,
+        moe_n_experts=best_params.get("moe_n_experts", moe_n_experts),
+        moe_top_k=best_params.get("moe_top_k", moe_top_k),
+        moe_expert_hidden_mult=best_params.get("moe_expert_hidden_mult", moe_expert_hidden_mult),
         moe_jitter_noise=moe_jitter_noise,
+        set_transformer_block=best_params.get("set_transformer_block", "sab"),
+        llm_kwargs={**llm_kwargs, **({"llm_use_lora": True, "llm_lora_r": best_params["llm_lora_r"]}
+                                     if (llm_kwargs.get("use_llm", False) and "llm_lora_r" in best_params) else {})},
     )
     model.rdkit_encoder._cache = rdkit_cache
 
@@ -676,7 +723,7 @@ def _run_single_outer_fold(
     train_result = train_fixed_epochs(
         model=model,
         train_loader=train_loader,
-        val_loader=None,
+        val_loader=monitor_loader,
         device=device,
         lr=best_params["lr"],
         weight_decay=best_params["weight_decay"],
@@ -684,6 +731,7 @@ def _run_single_outer_fold(
         class_weights=class_weights,
         use_cosine_annealing=True,
         backbone_lr_ratio=best_params.get("backbone_lr_ratio", 1.0),
+        freeze_backbone_epochs=3,
     )
 
     model.load_state_dict(train_result["final_state"])
@@ -695,6 +743,7 @@ def _run_single_outer_fold(
         "n_attn_layers": best_params["n_attn_layers"],
         "fusion_strategy": best_params["fusion_strategy"],
         "head_hidden_dim": best_params["head_hidden_dim"],
+        "set_transformer_block": best_params.get("set_transformer_block", "sab"),
         "dropout": best_params["dropout"],
         "use_mpnn": use_mpnn,
         "use_moe": use_moe,
@@ -702,6 +751,7 @@ def _run_single_outer_fold(
         "moe_top_k": moe_top_k,
         "moe_expert_hidden_mult": moe_expert_hidden_mult,
         "moe_jitter_noise": moe_jitter_noise,
+        **(llm_kwargs or {}),
     }
 
     torch.save({
@@ -759,12 +809,24 @@ def main(
     load_delivery_head: bool = False,
     # MPNN
     use_mpnn: bool = False,
-    # MoE（新增）
+    n_repeats: int = 1,
+    repeat_seed_step: int = 1000,
+    # MoE（消融开关）
     use_moe: bool = False,
     moe_n_experts: int = 4,
     moe_top_k: int = 2,
     moe_expert_hidden_mult: int = 2,
     moe_jitter_noise: float = 0.0,
+    # Set Transformer
+    set_transformer_block: str = "sab",
+    # LLM（消融开关）
+    use_llm: bool = False,
+    llm_model_path: str = DEFAULT_MOLT5_PATH,
+    llm_freeze: bool = True,
+    llm_use_lora: bool = False,
+    llm_lora_r: int = 8,
+    llm_lora_alpha: int = 16,
+    llm_lora_dropout: float = 0.05,
     # 并行
     parallel: bool = False,
     # 设备
@@ -785,6 +847,17 @@ def main(
     
     logger.info(f"Using device: {device}")
     device = torch.device(device)
+    set_global_seed(seed)
+
+    llm_kwargs = dict(
+        use_llm=use_llm,
+        llm_model_path=llm_model_path,
+        llm_freeze=llm_freeze,
+        llm_use_lora=llm_use_lora,
+        llm_lora_r=llm_lora_r,
+        llm_lora_alpha=llm_lora_alpha,
+        llm_lora_dropout=llm_lora_dropout,
+    )
     
     # 加载预训练权重（如果指定）
     pretrain_state_dict = None
@@ -858,6 +931,8 @@ def main(
             moe_top_k=moe_top_k,
             moe_expert_hidden_mult=moe_expert_hidden_mult,
             moe_jitter_noise=moe_jitter_noise,
+            set_transformer_block=set_transformer_block,
+            llm_kwargs=llm_kwargs,
         ))
     
     if parallel:
@@ -873,9 +948,22 @@ def main(
     else:
         logger.info(f"Running {n_outer_folds} outer folds SEQUENTIALLY")
         outer_results = []
+        per_fold_params = {}
         for args in fold_args:
             result = _run_single_outer_fold(**args)
             outer_results.append(result)
+            per_fold_params[result["fold"]] = (result["best_params"], result["epoch_mean"])
+
+        for rep in range(1, n_repeats):
+            logger.info(f"\n===== Repeat {rep}/{n_repeats - 1} (reuse hyperparams) =====")
+            for args in fold_args:
+                bp, em = per_fold_params[args["outer_fold"]]
+                rep_args = dict(args)
+                rep_args["seed"] = seed + rep * repeat_seed_step
+                rep_args["fold_dir"] = run_dir / f"repeat_{rep}" / f"outer_fold_{args['outer_fold']}"
+                rep_args["precomputed_best_params"] = bp
+                rep_args["precomputed_epoch_mean"] = em
+                outer_results.append(_run_single_outer_fold(**rep_args))
     
     # 汇总结果
     logger.info("\n" + "=" * 60)

lnp_ml/modeling/pretrain.py

@@ -31,6 +31,8 @@ def find_mpnn_ensemble_paths(base_dir: Path = DEFAULT_MPNN_ENSEMBLE_DIR) -> List
         raise FileNotFoundError(f"No model.pt files found in {base_dir}")
     return [str(p) for p in model_paths]
 from lnp_ml.modeling.models import LNPModel, LNPModelWithoutMPNN
+from lnp_ml.modeling.layers.llm_prompt import DEFAULT_MOLT5_PATH
+from lnp_ml.utils.seed import set_global_seed
 
 
 app = typer.Typer()
@@ -156,7 +158,8 @@ def pretrain(
         训练历史和最佳验证损失
     """
     model = model.to(device)
-    optimizer = torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=weight_decay)
+    trainable_params = [p for p in model.parameters() if p.requires_grad]
+    optimizer = torch.optim.AdamW(trainable_params, lr=lr, weight_decay=weight_decay)
     scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
         optimizer, mode="min", factor=0.5, patience=5
     )
@@ -217,13 +220,28 @@ def main(
     d_model: int = 256,
     num_heads: int = 8,
     n_attn_layers: int = 4,
+    set_transformer_block: str = "sab",  # "sab" | "isab"
     fusion_strategy: str = "attention",
     head_hidden_dim: int = 128,
     dropout: float = 0.1,
+    # MoE 参数（消融开关）
+    use_moe: bool = False,
+    moe_n_experts: int = 4,
+    moe_top_k: int = 2,
+    moe_expert_hidden_mult: int = 2,
+    moe_jitter_noise: float = 0.0,
+    # LLM 参数（消融开关）
+    use_llm: bool = False,
+    llm_model_path: str = DEFAULT_MOLT5_PATH,
+    llm_freeze: bool = True,
+    llm_use_lora: bool = False,
+    llm_lora_r: int = 8,
+    llm_lora_alpha: int = 16,
+    llm_lora_dropout: float = 0.05,
     # MPNN 参数（可选）
     use_mpnn: bool = False,  # 启用 MPNN，自动从默认路径加载 ensemble
     mpnn_checkpoint: Optional[str] = None,
-    mpnn_ensemble_paths: Optional[str] = None,  # 逗号分隔的路径列表
+    mpnn_ensemble_paths: Optional[str] = None,
     mpnn_device: str = "cpu",
     # 训练参数
     batch_size: int = 64,
@@ -231,6 +249,8 @@ def main(
     weight_decay: float = 1e-5,
     epochs: int = 50,
     patience: int = 10,
+    # 随机种子
+    seed: int = 42,
     # 设备
     device: str = "cuda" if torch.cuda.is_available() else "cpu",
 ):
@@ -245,7 +265,8 @@ def main(
         - models/pretrain_delivery.pt: 包含 backbone + delivery head 权重
         - models/pretrain_history.json: 训练历史
     """
-    logger.info(f"Using device: {device}")
+    set_global_seed(seed)
+    logger.info(f"Using device: {device} | seed: {seed}")
     device_obj = torch.device(device)
 
     # 加载已处理的 parquet 文件
@@ -280,28 +301,39 @@ def main(
     enable_mpnn = mpnn_checkpoint is not None or ensemble_paths_list is not None
 
     # 创建模型
-    logger.info(f"Creating model (use_mpnn={enable_mpnn})...")
+    logger.info(
+        f"Creating model (use_mpnn={enable_mpnn}, use_moe={use_moe}, use_llm={use_llm})..."
+    )
+    common_kwargs = dict(
+        d_model=d_model,
+        num_heads=num_heads,
+        n_attn_layers=n_attn_layers,
+        set_transformer_block=set_transformer_block,
+        fusion_strategy=fusion_strategy,
+        head_hidden_dim=head_hidden_dim,
+        dropout=dropout,
+        use_moe=use_moe,
+        moe_n_experts=moe_n_experts,
+        moe_top_k=moe_top_k,
+        moe_expert_hidden_mult=moe_expert_hidden_mult,
+        moe_jitter_noise=moe_jitter_noise,
+        use_llm=use_llm,
+        llm_model_path=llm_model_path,
+        llm_freeze=llm_freeze,
+        llm_use_lora=llm_use_lora,
+        llm_lora_r=llm_lora_r,
+        llm_lora_alpha=llm_lora_alpha,
+        llm_lora_dropout=llm_lora_dropout,
+    )
     if enable_mpnn:
         model = LNPModel(
-            d_model=d_model,
-            num_heads=num_heads,
-            n_attn_layers=n_attn_layers,
-            fusion_strategy=fusion_strategy,
-            head_hidden_dim=head_hidden_dim,
-            dropout=dropout,
             mpnn_checkpoint=mpnn_checkpoint,
             mpnn_ensemble_paths=ensemble_paths_list,
             mpnn_device=mpnn_device,
+            **common_kwargs,
         )
     else:
-        model = LNPModelWithoutMPNN(
-            d_model=d_model,
-            num_heads=num_heads,
-            n_attn_layers=n_attn_layers,
-            fusion_strategy=fusion_strategy,
-            head_hidden_dim=head_hidden_dim,
-            dropout=dropout,
-        )
+        model = LNPModelWithoutMPNN(**common_kwargs)
 
     n_params_total = sum(p.numel() for p in model.parameters())
     n_params_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
@@ -336,10 +368,23 @@ def main(
                 "d_model": d_model,
                 "num_heads": num_heads,
                 "n_attn_layers": n_attn_layers,
+                "set_transformer_block": set_transformer_block,
                 "fusion_strategy": fusion_strategy,
                 "head_hidden_dim": head_hidden_dim,
                 "dropout": dropout,
                 "use_mpnn": enable_mpnn,
+                "use_moe": use_moe,
+                "moe_n_experts": moe_n_experts,
+                "moe_top_k": moe_top_k,
+                "moe_expert_hidden_mult": moe_expert_hidden_mult,
+                "moe_jitter_noise": moe_jitter_noise,
+                "use_llm": use_llm,
+                "llm_model_path": llm_model_path,
+                "llm_freeze": llm_freeze,
+                "llm_use_lora": llm_use_lora,
+                "llm_lora_r": llm_lora_r,
+                "llm_lora_alpha": llm_lora_alpha,
+                "llm_lora_dropout": llm_lora_dropout,
             },
             "best_val_loss": result["best_val_loss"],
         },
@@ -398,32 +443,40 @@ def test(
     
     # 解析 MPNN 配置
     enable_mpnn = config.get("use_mpnn", False)
+    common_kwargs = dict(
+        d_model=config["d_model"],
+        num_heads=config["num_heads"],
+        n_attn_layers=config["n_attn_layers"],
+        set_transformer_block=config.get("set_transformer_block", "sab"),
+        fusion_strategy=config["fusion_strategy"],
+        head_hidden_dim=config["head_hidden_dim"],
+        dropout=config["dropout"],
+        use_moe=config.get("use_moe", False),
+        moe_n_experts=config.get("moe_n_experts", 4),
+        moe_top_k=config.get("moe_top_k", 2),
+        moe_expert_hidden_mult=config.get("moe_expert_hidden_mult", 2),
+        moe_jitter_noise=config.get("moe_jitter_noise", 0.0),
+        use_llm=config.get("use_llm", False),
+        llm_model_path=config.get("llm_model_path", DEFAULT_MOLT5_PATH),
+        llm_freeze=config.get("llm_freeze", True),
+        llm_use_lora=config.get("llm_use_lora", False),
+        llm_lora_r=config.get("llm_lora_r", 8),
+        llm_lora_alpha=config.get("llm_lora_alpha", 16),
+        llm_lora_dropout=config.get("llm_lora_dropout", 0.05),
+    )
     if enable_mpnn or use_mpnn:
         logger.info(f"Auto-detecting MPNN ensemble from {DEFAULT_MPNN_ENSEMBLE_DIR}")
         ensemble_paths = find_mpnn_ensemble_paths()
         logger.info(f"Found {len(ensemble_paths)} MPNN models")
-        
         model = LNPModel(
-            d_model=config["d_model"],
-            num_heads=config["num_heads"],
-            n_attn_layers=config["n_attn_layers"],
-            fusion_strategy=config["fusion_strategy"],
-            head_hidden_dim=config["head_hidden_dim"],
-            dropout=config["dropout"],
             mpnn_ensemble_paths=ensemble_paths,
             mpnn_device=mpnn_device,
+            **common_kwargs,
         )
     else:
-        model = LNPModelWithoutMPNN(
-            d_model=config["d_model"],
-            num_heads=config["num_heads"],
-            n_attn_layers=config["n_attn_layers"],
-            fusion_strategy=config["fusion_strategy"],
-            head_hidden_dim=config["head_hidden_dim"],
-            dropout=config["dropout"],
-        )
+        model = LNPModelWithoutMPNN(**common_kwargs)
 
-    model.load_state_dict(checkpoint["model_state_dict"])
+    model.load_state_dict(checkpoint["model_state_dict"], strict=False)
     model.to(device_obj)
     model.eval()
     

lnp_ml/modeling/trainer_balanced.py

@@ -29,6 +29,7 @@ class LossWeightsBalanced:
     biodist: float = 1.0
     toxic: float = 1.0
     moe_lb: float = 0.01  # MoE load-balancing 系数（仅在 use_moe=True 时生效）
+    moe_lb: float = 0.01  # MoE load-balancing 系数（仅在 use_moe=True 时生效）
 
 
 def compute_class_weights_from_loader(
@@ -112,6 +113,7 @@ def compute_multitask_loss_balanced(
     task_weights: Optional[LossWeightsBalanced] = None,
     class_weights: Optional[ClassWeights] = None,
     model: Optional[nn.Module] = None,
+    model: Optional[nn.Module] = None,
 ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
     """
     计算带类权重的多任务损失。
@@ -129,66 +131,59 @@ def compute_multitask_loss_balanced(
     task_weights = task_weights or LossWeightsBalanced()
     class_weights = class_weights or ClassWeights()
 
+    # 若模型 head 暴露 log_vars，则启用不确定性加权（自动平衡），否则回退到标量权重
+    log_vars = None
+    if model is not None and hasattr(model, "head") and hasattr(model.head, "log_vars"):
+        log_vars = model.head.log_vars
+
     losses = {}
     device = next(iter(outputs.values())).device
     total_loss = torch.tensor(0.0, device=device)
 
-    # size: MSE loss（回归任务，不需要类权重）
+    def _add(name: str, raw_loss: torch.Tensor, scalar_w: float) -> None:
+        nonlocal total_loss
+        losses[name] = raw_loss
+        if log_vars is not None and name in log_vars:
+            s = log_vars[name]
+            total_loss = total_loss + torch.exp(-s) * raw_loss + 0.5 * s
+        else:
+            total_loss = total_loss + scalar_w * raw_loss
+
+    # size: MSE
     if "size" in targets and mask["size"].any():
         m = mask["size"]
-        pred = outputs["size"][m].squeeze(-1)
-        tgt = targets["size"][m]
-        losses["size"] = F.mse_loss(pred, tgt)
-        total_loss = total_loss + task_weights.size * losses["size"]
+        _add("size", F.mse_loss(outputs["size"][m].squeeze(-1), targets["size"][m]), task_weights.size)
 
-    # delivery: MSE loss（回归任务，不需要类权重）
+    # delivery: MSE
     if "delivery" in targets and mask["delivery"].any():
         m = mask["delivery"]
-        pred = outputs["delivery"][m].squeeze(-1)
-        tgt = targets["delivery"][m]
-        losses["delivery"] = F.mse_loss(pred, tgt)
-        total_loss = total_loss + task_weights.delivery * losses["delivery"]
+        _add("delivery", F.mse_loss(outputs["delivery"][m].squeeze(-1), targets["delivery"][m]), task_weights.delivery)
 
-    # pdi: CrossEntropy with class weights
+    # pdi: 类加权 CE
     if "pdi" in targets and mask["pdi"].any():
         m = mask["pdi"]
-        pred = outputs["pdi"][m]
-        tgt = targets["pdi"][m]
-        weight = class_weights.pdi.to(device) if class_weights.pdi is not None else None
-        losses["pdi"] = F.cross_entropy(pred, tgt, weight=weight)
-        total_loss = total_loss + task_weights.pdi * losses["pdi"]
+        w = class_weights.pdi.to(device) if class_weights.pdi is not None else None
+        _add("pdi", F.cross_entropy(outputs["pdi"][m], targets["pdi"][m], weight=w), task_weights.pdi)
 
-    # ee: CrossEntropy with class weights
+    # ee: 类加权 CE
     if "ee" in targets and mask["ee"].any():
         m = mask["ee"]
-        pred = outputs["ee"][m]
-        tgt = targets["ee"][m]
-        weight = class_weights.ee.to(device) if class_weights.ee is not None else None
-        losses["ee"] = F.cross_entropy(pred, tgt, weight=weight)
-        total_loss = total_loss + task_weights.ee * losses["ee"]
+        w = class_weights.ee.to(device) if class_weights.ee is not None else None
+        _add("ee", F.cross_entropy(outputs["ee"][m], targets["ee"][m], weight=w), task_weights.ee)
 
-    # toxic: CrossEntropy with class weights
+    # toxic: 类加权 CE
     if "toxic" in targets and mask["toxic"].any():
         m = mask["toxic"]
-        pred = outputs["toxic"][m]
-        tgt = targets["toxic"][m]
-        weight = class_weights.toxic.to(device) if class_weights.toxic is not None else None
-        losses["toxic"] = F.cross_entropy(pred, tgt, weight=weight)
-        total_loss = total_loss + task_weights.toxic * losses["toxic"]
+        w = class_weights.toxic.to(device) if class_weights.toxic is not None else None
+        _add("toxic", F.cross_entropy(outputs["toxic"][m], targets["toxic"][m], weight=w), task_weights.toxic)
 
-    # biodist: KL divergence（分布任务，不需要类权重）
+    # biodist: KL
     if "biodist" in targets and mask["biodist"].any():
         m = mask["biodist"]
-        pred = outputs["biodist"][m]
-        tgt = targets["biodist"][m]
-        losses["biodist"] = F.kl_div(
-            pred.log().clamp(min=-100),
-            tgt,
-            reduction="batchmean",
-        )
-        total_loss = total_loss + task_weights.biodist * losses["biodist"]
+        kl = F.kl_div(outputs["biodist"][m].log().clamp(min=-100), targets["biodist"][m], reduction="batchmean")
+        _add("biodist", kl, task_weights.biodist)
 
-    # MoE load-balancing aux loss（可选）
+    # MoE load-balancing aux loss（固定小权重，不纳入不确定性加权）
     if model is not None and hasattr(model, "get_last_moe_extras"):
         extras = model.get_last_moe_extras()
         if extras is not None and "lb_loss" in extras:
@@ -210,6 +205,7 @@ def train_epoch_balanced(
     model.train()
     total_loss = 0.0
     task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic", "moe_lb"]}
+    task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic", "moe_lb"]}
     n_batches = 0
     
     for batch in tqdm(loader, desc="Training", leave=False):
@@ -222,6 +218,7 @@ def train_epoch_balanced(
         outputs = model(smiles, tabular)
         loss, losses = compute_multitask_loss_balanced(
             outputs, targets, mask, task_weights, class_weights, model=model,
+            outputs, targets, mask, task_weights, class_weights, model=model,
         )
         
         loss.backward()
@@ -251,6 +248,7 @@ def validate_balanced(
     model.eval()
     total_loss = 0.0
     task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic", "moe_lb"]}
+    task_losses = {k: 0.0 for k in ["size", "pdi", "ee", "delivery", "biodist", "toxic", "moe_lb"]}
     n_batches = 0
     
     # 用于计算准确率
@@ -266,6 +264,7 @@ def validate_balanced(
         outputs = model(smiles, tabular)
         loss, losses = compute_multitask_loss_balanced(
             outputs, targets, mask, task_weights, class_weights, model=model,
+            outputs, targets, mask, task_weights, class_weights, model=model,
         )
         
         total_loss += loss.item()
@@ -295,7 +294,20 @@ def validate_balanced(
     return metrics
 
 
-BACKBONE_PREFIXES = ("token_projector.", "cross_attention.", "fusion.", "moe.")
+BACKBONE_PREFIXES = (
+    "token_projector.",
+    "set_transformer.",
+    "fusion.",
+    "moe.",
+    "llm_prompt.",
+)
+
+FROM_SCRATCH_PREFIXES = (
+    "moe.",
+    "llm_prompt.",
+    "fusion.g_moe",
+    "fusion.g_llm",
+)
 
 
 def build_optimizer(
@@ -307,18 +319,21 @@ def build_optimizer(
     """
     构建 AdamW 优化器，支持分层学习率。
 
+    仅收集 requires_grad=True 的参数（冻结的 MolT5 encoder 被排除）。
     当 backbone_lr_ratio < 1.0 时，backbone 参数使用 lr * backbone_lr_ratio，
     其余参数（task heads 等）使用 lr。backbone_lr_ratio = 1.0 等价于统一学习率。
     """
+    trainable = [(n, p) for n, p in model.named_parameters() if p.requires_grad]
+
     if backbone_lr_ratio >= 1.0:
-        return torch.optim.AdamW(model.parameters(), lr=lr, weight_decay=weight_decay)
+        return torch.optim.AdamW(
+            [p for _, p in trainable], lr=lr, weight_decay=weight_decay
+        )
 
     backbone_params = []
     head_params = []
-    for name, param in model.named_parameters():
-        if not param.requires_grad:
-            continue
-        if name.startswith(BACKBONE_PREFIXES):
+    for name, param in trainable:
+        if name.startswith(BACKBONE_PREFIXES) and not name.startswith(FROM_SCRATCH_PREFIXES):
             backbone_params.append(param)
         else:
             head_params.append(param)
@@ -442,6 +457,7 @@ def train_fixed_epochs(
     use_swa: bool = False,
     swa_start_epoch: Optional[int] = None,
     backbone_lr_ratio: float = 1.0,
+    freeze_backbone_epochs: int = 0,
 ) -> Dict:
     """
     固定 epoch 数的训练（不使用 early stopping）。
@@ -469,6 +485,15 @@ def train_fixed_epochs(
     model = model.to(device)
     optimizer = build_optimizer(model, lr, weight_decay, backbone_lr_ratio)
 
+    backbone_named = [
+        (n, p) for n, p in model.named_parameters()
+        if n.startswith(BACKBONE_PREFIXES) and not n.startswith(FROM_SCRATCH_PREFIXES)
+        and p.requires_grad
+    ]
+    if freeze_backbone_epochs > 0:
+        for _, p in backbone_named:
+            p.requires_grad_(False)
+
     if use_cosine_annealing:
         scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=epochs)
     else:
@@ -486,6 +511,9 @@ def train_fixed_epochs(
     history = {"train": [], "val": []}
     
     for epoch in range(epochs):
+        if freeze_backbone_epochs > 0 and epoch == freeze_backbone_epochs:
+            for _, p in backbone_named:
+                p.requires_grad_(True)
         # Train
         train_metrics = train_epoch_balanced(
             model, train_loader, optimizer, device, task_weights, class_weights

lnp_ml/utils/seed.py

@@ -0,0 +1,19 @@
+"""全局随机种子工具。"""
+
+import random
+
+import numpy as np
+import torch
+
+
+def set_global_seed(seed: int) -> None:
+    """固定 random / numpy / torch 种子，提升可复现性。
+
+    不开启 cudnn deterministic，以免显著拖慢训练；如需严格复现可自行追加：
+        torch.backends.cudnn.deterministic = True
+        torch.backends.cudnn.benchmark = False
+    """
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)

models/abl/baseline/20260617_180556/outer_fold_0/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.28242799368681437,
+  "lr": 0.00037183641805732076,
+  "weight_decay": 6.290644294586152e-05,
+  "backbone_lr_ratio": 0.10677482709481352,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/outer_fold_0/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 19, 21, 22, 23, 24, 25, 26, 27, 28, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 52, 53, 54, 55, 56, 57, 58, 61, 62, 63, 64, 65, 66, 67, 68, 71, 72, 73, 75, 76, 77, 80, 81, 85, 86, 88, 89, 90, 92, 94, 95, 96, 98, 99, 100, 101, 103, 104, 105, 106, 107, 108, 109, 111, 113, 114, 115, 116, 117, 118, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 133, 134, 135, 136, 138, 139, 141, 144, 145, 146, 147, 148, 149, 150, 151, 153, 154, 155, 156, 157, 158, 159, 160, 163, 164, 167, 168, 169, 170, 171, 172, 175, 176, 178, 179, 180, 182, 183, 184, 185, 188, 189, 190, 191, 192, 193, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 206, 208, 210, 211, 212, 213, 216, 217, 218, 219, 220, 222, 224, 225, 227, 228, 229, 230, 231, 232, 233, 234, 236, 237, 239, 241, 242, 244, 245, 247, 248, 249, 250, 252, 253, 255, 256, 257, 258, 259, 260, 261, 262, 265, 267, 268, 269, 270, 271, 272, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 285, 286, 287, 288, 291, 292, 294, 296, 297, 298, 299, 300, 301, 303, 304, 306, 307, 308, 309, 310, 311, 312, 313, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329, 330, 332, 333, 335, 336, 337, 338, 339, 340, 342, 343, 344, 345, 346, 347, 348, 349, 351, 352, 354, 355, 356, 357, 358, 359, 360, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 374, 375, 377, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389, 390, 391, 393, 395, 396, 397, 398, 399, 400, 401, 403, 404, 405, 406, 407, 408, 409, 411, 412, 413, 414, 415, 417, 418, 419], "outer_test_idx": [7, 16, 20, 29, 51, 59, 60, 69, 70, 74, 78, 79, 82, 83, 84, 87, 91, 93, 97, 102, 110, 112, 119, 120, 121, 132, 137, 140, 142, 143, 152, 161, 162, 165, 166, 173, 174, 177, 181, 186, 187, 194, 205, 207, 209, 214, 215, 221, 223, 226, 235, 238, 240, 243, 246, 251, 254, 263, 264, 266, 273, 284, 289, 290, 293, 295, 302, 305, 314, 315, 331, 334, 341, 350, 353, 361, 373, 376, 378, 392, 394, 402, 410, 416]}

models/abl/baseline/20260617_180556/outer_fold_0/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.32790454280151365,
+    "rmse": 0.572629498717551,
+    "mae": 0.35871042688208893,
+    "r2": -0.08876073797082196
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.7250728333301815,
+    "rmse": 0.8515120864263651,
+    "mae": 0.7098816664696768,
+    "r2": 0.07838867510225656
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5476190476190477,
+    "precision": 0.3435971685971686,
+    "recall": 0.5359094457455114,
+    "f1": 0.34977324263038545
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.7142857142857143,
+    "precision": 0.6871101871101871,
+    "recall": 0.707142857142857,
+    "f1": 0.6756066411238826
+  },
+  "toxic": {
+    "n_samples": 58,
+    "accuracy": 0.9655172413793104,
+    "precision": 0.75,
+    "recall": 0.9821428571428572,
+    "f1": 0.8242424242424242
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.4073394583325288,
+    "js_divergence": 0.092887269703335
+  }
+}

models/abl/baseline/20260617_180556/outer_fold_1/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.23085562232445592,
+  "lr": 0.0005227270589225132,
+  "weight_decay": 0.00461701039547356,
+  "backbone_lr_ratio": 0.12087164274488635,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/outer_fold_1/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 3, 5, 6, 7, 10, 11, 12, 13, 15, 16, 18, 19, 20, 21, 24, 26, 27, 29, 30, 31, 32, 33, 35, 36, 38, 39, 40, 42, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 65, 66, 67, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 97, 98, 99, 100, 101, 102, 103, 105, 106, 107, 108, 109, 110, 112, 114, 115, 118, 119, 120, 121, 123, 124, 125, 126, 128, 129, 130, 131, 132, 136, 137, 140, 141, 142, 143, 144, 145, 146, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 161, 162, 164, 165, 166, 167, 168, 169, 171, 172, 173, 174, 175, 176, 177, 178, 179, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 194, 195, 196, 197, 198, 199, 201, 202, 204, 205, 206, 207, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 225, 226, 228, 229, 231, 232, 233, 234, 235, 236, 237, 238, 240, 242, 243, 244, 246, 247, 248, 250, 251, 252, 254, 255, 256, 258, 259, 260, 261, 262, 263, 264, 266, 268, 269, 270, 271, 273, 274, 275, 276, 277, 278, 279, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 295, 296, 297, 300, 301, 302, 303, 304, 305, 306, 307, 308, 310, 311, 313, 314, 315, 316, 317, 318, 319, 321, 322, 323, 324, 325, 327, 328, 329, 330, 331, 333, 334, 335, 337, 338, 339, 340, 341, 342, 343, 344, 345, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 361, 362, 364, 365, 366, 367, 368, 369, 373, 374, 375, 376, 378, 379, 380, 381, 382, 384, 387, 388, 389, 391, 392, 393, 394, 396, 397, 398, 399, 402, 403, 404, 405, 406, 408, 409, 410, 411, 412, 413, 414, 416, 417, 419], "outer_test_idx": [1, 2, 4, 8, 9, 14, 17, 22, 23, 25, 28, 34, 37, 41, 43, 45, 63, 64, 68, 85, 96, 104, 111, 113, 116, 117, 122, 127, 133, 134, 135, 138, 139, 147, 160, 163, 170, 180, 191, 192, 193, 200, 203, 208, 224, 227, 230, 239, 241, 245, 249, 253, 257, 265, 267, 272, 280, 294, 298, 299, 309, 312, 320, 326, 332, 336, 346, 347, 360, 363, 370, 371, 372, 377, 383, 385, 386, 390, 395, 400, 401, 407, 415, 418]}

models/abl/baseline/20260617_180556/outer_fold_1/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.233876145855699,
+    "rmse": 0.4836074294876982,
+    "mae": 0.3760785318556286,
+    "r2": -2.874543407510119
+  },
+  "delivery": {
+    "n_samples": 61,
+    "mse": 1.0238964615450643,
+    "rmse": 1.0118776910007772,
+    "mae": 0.6578027546466862,
+    "r2": 0.2249841902125319
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5238095238095238,
+    "precision": 0.34497354497354493,
+    "recall": 0.3280423280423281,
+    "f1": 0.3231922398589065
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.5595238095238095,
+    "precision": 0.5101731601731602,
+    "recall": 0.5036154321868608,
+    "f1": 0.48790324563862697
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9508196721311475,
+    "precision": 0.75,
+    "recall": 0.9741379310344828,
+    "f1": 0.8200589970501475
+  },
+  "biodist": {
+    "n_samples": 61,
+    "kl_divergence": 0.48528378254405913,
+    "js_divergence": 0.12258377401744426
+  }
+}

models/abl/baseline/20260617_180556/outer_fold_2/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.18117503720006686,
+  "lr": 8.135394049241399e-05,
+  "weight_decay": 0.015076469897649124,
+  "backbone_lr_ratio": 0.9586175198423679,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/outer_fold_2/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 11, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 32, 33, 34, 36, 37, 38, 39, 40, 41, 43, 44, 45, 47, 48, 49, 50, 51, 55, 56, 57, 58, 59, 60, 61, 63, 64, 65, 68, 69, 70, 72, 74, 76, 78, 79, 80, 82, 83, 84, 85, 86, 87, 88, 91, 93, 94, 95, 96, 97, 100, 101, 102, 103, 104, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 142, 143, 145, 147, 148, 149, 150, 151, 152, 153, 156, 157, 158, 159, 160, 161, 162, 163, 165, 166, 167, 168, 170, 171, 172, 173, 174, 175, 177, 179, 180, 181, 183, 184, 185, 186, 187, 188, 191, 192, 193, 194, 196, 197, 199, 200, 202, 203, 204, 205, 207, 208, 209, 210, 211, 212, 213, 214, 215, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 232, 233, 235, 236, 237, 238, 239, 240, 241, 242, 243, 245, 246, 249, 250, 251, 252, 253, 254, 255, 256, 257, 259, 260, 261, 263, 264, 265, 266, 267, 268, 269, 271, 272, 273, 274, 275, 276, 280, 281, 282, 283, 284, 285, 287, 289, 290, 293, 294, 295, 296, 297, 298, 299, 300, 302, 303, 305, 307, 308, 309, 310, 311, 312, 313, 314, 315, 317, 318, 320, 321, 322, 323, 324, 326, 327, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 358, 360, 361, 363, 364, 365, 366, 368, 369, 370, 371, 372, 373, 376, 377, 378, 379, 380, 381, 382, 383, 385, 386, 387, 388, 389, 390, 392, 394, 395, 398, 399, 400, 401, 402, 404, 405, 407, 408, 409, 410, 412, 413, 414, 415, 416, 417, 418], "outer_test_idx": [10, 12, 15, 31, 35, 42, 46, 52, 53, 54, 62, 66, 67, 71, 73, 75, 77, 81, 89, 90, 92, 98, 99, 105, 123, 124, 141, 144, 146, 154, 155, 164, 169, 176, 178, 182, 189, 190, 195, 198, 201, 206, 216, 217, 218, 219, 231, 234, 244, 247, 248, 258, 262, 270, 277, 278, 279, 286, 288, 291, 292, 301, 304, 306, 316, 319, 325, 328, 343, 357, 359, 362, 367, 374, 375, 384, 391, 393, 396, 397, 403, 406, 411, 419]}

models/abl/baseline/20260617_180556/outer_fold_2/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.07870736214858826,
+    "rmse": 0.28054832408800495,
+    "mae": 0.2252055633635748,
+    "r2": 0.15667568332606574
+  },
+  "delivery": {
+    "n_samples": 60,
+    "mse": 0.6624288051399898,
+    "rmse": 0.8138972939750997,
+    "mae": 0.5578871982870623,
+    "r2": 0.118192445613993
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5714285714285714,
+    "precision": 0.43453195231581004,
+    "recall": 0.7014848950332823,
+    "f1": 0.4452380952380952
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6666666666666666,
+    "precision": 0.6140109890109889,
+    "recall": 0.6785714285714285,
+    "f1": 0.6262626262626263
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9344262295081968,
+    "precision": 0.7142857142857143,
+    "recall": 0.9655172413793103,
+    "f1": 0.7821428571428571
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.3967825471495568,
+    "js_divergence": 0.1057568924841401
+  }
+}

models/abl/baseline/20260617_180556/outer_fold_3/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.1799936812783694,
+  "lr": 0.00033865324369584493,
+  "weight_decay": 0.01440133094015265,
+  "backbone_lr_ratio": 0.3278334190266755,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/outer_fold_3/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 4, 5, 7, 8, 9, 10, 11, 12, 14, 15, 16, 17, 20, 21, 22, 23, 24, 25, 27, 28, 29, 30, 31, 32, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 45, 46, 47, 51, 52, 53, 54, 55, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 73, 74, 75, 76, 77, 78, 79, 81, 82, 83, 84, 85, 86, 87, 89, 90, 91, 92, 93, 95, 96, 97, 98, 99, 100, 102, 104, 105, 106, 109, 110, 111, 112, 113, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 127, 129, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 146, 147, 148, 150, 151, 152, 153, 154, 155, 156, 158, 159, 160, 161, 162, 163, 164, 165, 166, 168, 169, 170, 172, 173, 174, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 189, 190, 191, 192, 193, 194, 195, 198, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 212, 213, 214, 215, 216, 217, 218, 219, 221, 223, 224, 226, 227, 228, 230, 231, 233, 234, 235, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 251, 252, 253, 254, 255, 256, 257, 258, 259, 262, 263, 264, 265, 266, 267, 268, 269, 270, 272, 273, 274, 275, 276, 277, 278, 279, 280, 282, 284, 286, 288, 289, 290, 291, 292, 293, 294, 295, 297, 298, 299, 301, 302, 304, 305, 306, 307, 308, 309, 310, 312, 314, 315, 316, 317, 319, 320, 321, 325, 326, 327, 328, 331, 332, 334, 336, 340, 341, 343, 344, 346, 347, 348, 350, 352, 353, 357, 359, 360, 361, 362, 363, 364, 366, 367, 368, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 382, 383, 384, 385, 386, 387, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 400, 401, 402, 403, 404, 406, 407, 410, 411, 412, 414, 415, 416, 418, 419], "outer_test_idx": [3, 6, 13, 18, 19, 26, 33, 44, 48, 49, 50, 56, 57, 72, 80, 88, 94, 101, 103, 107, 108, 114, 126, 128, 130, 145, 149, 157, 167, 171, 175, 188, 196, 197, 199, 211, 220, 222, 225, 229, 232, 236, 250, 260, 261, 271, 281, 283, 285, 287, 296, 300, 303, 311, 313, 318, 322, 323, 324, 329, 330, 333, 335, 337, 338, 339, 342, 345, 349, 351, 354, 355, 356, 358, 365, 369, 381, 388, 399, 405, 408, 409, 413, 417]}

models/abl/baseline/20260617_180556/outer_fold_3/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.477705869562307,
+    "rmse": 0.6911626939891266,
+    "mae": 0.38659381004701177,
+    "r2": -0.46585807501413723
+  },
+  "delivery": {
+    "n_samples": 59,
+    "mse": 0.7692103143544243,
+    "rmse": 0.8770463581558413,
+    "mae": 0.6155587352528158,
+    "r2": 0.2522361838732067
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.6904761904761905,
+    "precision": 0.4426767676767677,
+    "recall": 0.6817716333845366,
+    "f1": 0.43613724611708476
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6428571428571429,
+    "precision": 0.5650987224157956,
+    "recall": 0.5790095525389644,
+    "f1": 0.5637508626639062
+  },
+  "toxic": {
+    "n_samples": 60,
+    "accuracy": 0.9333333333333333,
+    "precision": 0.7142857142857143,
+    "recall": 0.9649122807017544,
+    "f1": 0.7818181818181817
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.4102080974965646,
+    "js_divergence": 0.11100613154873733
+  }
+}

models/abl/baseline/20260617_180556/outer_fold_4/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.14162975892531532,
+  "lr": 0.0004019907813153378,
+  "weight_decay": 0.00037960139235180325,
+  "backbone_lr_ratio": 0.18010931581670617,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/outer_fold_4/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [1, 2, 3, 4, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17, 18, 19, 20, 22, 23, 25, 26, 28, 29, 31, 33, 34, 35, 37, 41, 42, 43, 44, 45, 46, 48, 49, 50, 51, 52, 53, 54, 56, 57, 59, 60, 62, 63, 64, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 77, 78, 79, 80, 81, 82, 83, 84, 85, 87, 88, 89, 90, 91, 92, 93, 94, 96, 97, 98, 99, 101, 102, 103, 104, 105, 107, 108, 110, 111, 112, 113, 114, 116, 117, 119, 120, 121, 122, 123, 124, 126, 127, 128, 130, 132, 133, 134, 135, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 149, 152, 154, 155, 157, 160, 161, 162, 163, 164, 165, 166, 167, 169, 170, 171, 173, 174, 175, 176, 177, 178, 180, 181, 182, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 205, 206, 207, 208, 209, 211, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 234, 235, 236, 238, 239, 240, 241, 243, 244, 245, 246, 247, 248, 249, 250, 251, 253, 254, 257, 258, 260, 261, 262, 263, 264, 265, 266, 267, 270, 271, 272, 273, 277, 278, 279, 280, 281, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 298, 299, 300, 301, 302, 303, 304, 305, 306, 309, 311, 312, 313, 314, 315, 316, 318, 319, 320, 322, 323, 324, 325, 326, 328, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 341, 342, 343, 345, 346, 347, 349, 350, 351, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 365, 367, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 381, 383, 384, 385, 386, 388, 390, 391, 392, 393, 394, 395, 396, 397, 399, 400, 401, 402, 403, 405, 406, 407, 408, 409, 410, 411, 413, 415, 416, 417, 418, 419], "outer_test_idx": [0, 5, 11, 21, 24, 27, 30, 32, 36, 38, 39, 40, 47, 55, 58, 61, 65, 76, 86, 95, 100, 106, 109, 115, 118, 125, 129, 131, 136, 148, 150, 151, 153, 156, 158, 159, 168, 172, 179, 183, 184, 185, 202, 204, 210, 212, 213, 228, 233, 237, 242, 252, 255, 256, 259, 268, 269, 274, 275, 276, 282, 297, 307, 308, 310, 317, 321, 327, 340, 344, 348, 352, 364, 366, 368, 379, 380, 382, 387, 389, 398, 404, 412, 414]}

models/abl/baseline/20260617_180556/outer_fold_4/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.1842198466764517,
+    "rmse": 0.4292083953937198,
+    "mae": 0.29178874833243235,
+    "r2": -0.33736192476816806
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.9003281069383875,
+    "rmse": 0.9488562098328637,
+    "mae": 0.6070586243837044,
+    "r2": 0.1293975832987566
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5476190476190477,
+    "precision": 0.34548450889914306,
+    "recall": 0.5337941628264209,
+    "f1": 0.3497882046269143
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.7023809523809523,
+    "precision": 0.6548765244417418,
+    "recall": 0.7576923076923077,
+    "f1": 0.6777508856852221
+  },
+  "toxic": {
+    "n_samples": 59,
+    "accuracy": 0.9661016949152542,
+    "precision": 0.8,
+    "recall": 0.9821428571428572,
+    "f1": 0.8659090909090909
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.3946735882065513,
+    "js_divergence": 0.1067245125767987
+  }
+}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_0/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.28242799368681437,
+  "lr": 0.00037183641805732076,
+  "weight_decay": 6.290644294586152e-05,
+  "backbone_lr_ratio": 0.10677482709481352,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_0/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 19, 21, 22, 23, 24, 25, 26, 27, 28, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 52, 53, 54, 55, 56, 57, 58, 61, 62, 63, 64, 65, 66, 67, 68, 71, 72, 73, 75, 76, 77, 80, 81, 85, 86, 88, 89, 90, 92, 94, 95, 96, 98, 99, 100, 101, 103, 104, 105, 106, 107, 108, 109, 111, 113, 114, 115, 116, 117, 118, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 133, 134, 135, 136, 138, 139, 141, 144, 145, 146, 147, 148, 149, 150, 151, 153, 154, 155, 156, 157, 158, 159, 160, 163, 164, 167, 168, 169, 170, 171, 172, 175, 176, 178, 179, 180, 182, 183, 184, 185, 188, 189, 190, 191, 192, 193, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 206, 208, 210, 211, 212, 213, 216, 217, 218, 219, 220, 222, 224, 225, 227, 228, 229, 230, 231, 232, 233, 234, 236, 237, 239, 241, 242, 244, 245, 247, 248, 249, 250, 252, 253, 255, 256, 257, 258, 259, 260, 261, 262, 265, 267, 268, 269, 270, 271, 272, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 285, 286, 287, 288, 291, 292, 294, 296, 297, 298, 299, 300, 301, 303, 304, 306, 307, 308, 309, 310, 311, 312, 313, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329, 330, 332, 333, 335, 336, 337, 338, 339, 340, 342, 343, 344, 345, 346, 347, 348, 349, 351, 352, 354, 355, 356, 357, 358, 359, 360, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 374, 375, 377, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389, 390, 391, 393, 395, 396, 397, 398, 399, 400, 401, 403, 404, 405, 406, 407, 408, 409, 411, 412, 413, 414, 415, 417, 418, 419], "outer_test_idx": [7, 16, 20, 29, 51, 59, 60, 69, 70, 74, 78, 79, 82, 83, 84, 87, 91, 93, 97, 102, 110, 112, 119, 120, 121, 132, 137, 140, 142, 143, 152, 161, 162, 165, 166, 173, 174, 177, 181, 186, 187, 194, 205, 207, 209, 214, 215, 221, 223, 226, 235, 238, 240, 243, 246, 251, 254, 263, 264, 266, 273, 284, 289, 290, 293, 295, 302, 305, 314, 315, 331, 334, 341, 350, 353, 361, 373, 376, 378, 392, 394, 402, 410, 416]}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_0/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.3808128541133737,
+    "rmse": 0.6171003598389598,
+    "mae": 0.4168719153806388,
+    "r2": -0.2644353156285024
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.751610048614176,
+    "rmse": 0.8669544674400012,
+    "mae": 0.7122741444999802,
+    "r2": 0.04465827311677528
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5,
+    "precision": 0.3181788931788932,
+    "recall": 0.5039032006245121,
+    "f1": 0.32369614512471656
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6666666666666666,
+    "precision": 0.6258687258687259,
+    "recall": 0.6461904761904762,
+    "f1": 0.6127803231251508
+  },
+  "toxic": {
+    "n_samples": 58,
+    "accuracy": 0.9655172413793104,
+    "precision": 0.75,
+    "recall": 0.9821428571428572,
+    "f1": 0.8242424242424242
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.39644445917049376,
+    "js_divergence": 0.09334798400608228
+  }
+}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_1/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.23085562232445592,
+  "lr": 0.0005227270589225132,
+  "weight_decay": 0.00461701039547356,
+  "backbone_lr_ratio": 0.12087164274488635,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_1/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 3, 5, 6, 7, 10, 11, 12, 13, 15, 16, 18, 19, 20, 21, 24, 26, 27, 29, 30, 31, 32, 33, 35, 36, 38, 39, 40, 42, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 65, 66, 67, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 97, 98, 99, 100, 101, 102, 103, 105, 106, 107, 108, 109, 110, 112, 114, 115, 118, 119, 120, 121, 123, 124, 125, 126, 128, 129, 130, 131, 132, 136, 137, 140, 141, 142, 143, 144, 145, 146, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 161, 162, 164, 165, 166, 167, 168, 169, 171, 172, 173, 174, 175, 176, 177, 178, 179, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 194, 195, 196, 197, 198, 199, 201, 202, 204, 205, 206, 207, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 225, 226, 228, 229, 231, 232, 233, 234, 235, 236, 237, 238, 240, 242, 243, 244, 246, 247, 248, 250, 251, 252, 254, 255, 256, 258, 259, 260, 261, 262, 263, 264, 266, 268, 269, 270, 271, 273, 274, 275, 276, 277, 278, 279, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 295, 296, 297, 300, 301, 302, 303, 304, 305, 306, 307, 308, 310, 311, 313, 314, 315, 316, 317, 318, 319, 321, 322, 323, 324, 325, 327, 328, 329, 330, 331, 333, 334, 335, 337, 338, 339, 340, 341, 342, 343, 344, 345, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 361, 362, 364, 365, 366, 367, 368, 369, 373, 374, 375, 376, 378, 379, 380, 381, 382, 384, 387, 388, 389, 391, 392, 393, 394, 396, 397, 398, 399, 402, 403, 404, 405, 406, 408, 409, 410, 411, 412, 413, 414, 416, 417, 419], "outer_test_idx": [1, 2, 4, 8, 9, 14, 17, 22, 23, 25, 28, 34, 37, 41, 43, 45, 63, 64, 68, 85, 96, 104, 111, 113, 116, 117, 122, 127, 133, 134, 135, 138, 139, 147, 160, 163, 170, 180, 191, 192, 193, 200, 203, 208, 224, 227, 230, 239, 241, 245, 249, 253, 257, 265, 267, 272, 280, 294, 298, 299, 309, 312, 320, 326, 332, 336, 346, 347, 360, 363, 370, 371, 372, 377, 383, 385, 386, 390, 395, 400, 401, 407, 415, 418]}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_1/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.3074589292084139,
+    "rmse": 0.5544897917982025,
+    "mae": 0.3793355481965201,
+    "r2": -4.093563359726248
+  },
+  "delivery": {
+    "n_samples": 61,
+    "mse": 1.0592195690639976,
+    "rmse": 1.029183933543464,
+    "mae": 0.6786946221025753,
+    "r2": 0.19824714422578682
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5952380952380952,
+    "precision": 0.37142213207121927,
+    "recall": 0.35978835978835977,
+    "f1": 0.35976608187134507
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.5476190476190477,
+    "precision": 0.5052287581699346,
+    "recall": 0.5094429380143666,
+    "f1": 0.4903811048389362
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9508196721311475,
+    "precision": 0.75,
+    "recall": 0.9741379310344828,
+    "f1": 0.8200589970501475
+  },
+  "biodist": {
+    "n_samples": 61,
+    "kl_divergence": 0.5219711265981701,
+    "js_divergence": 0.13035793372268764
+  }
+}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_2/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.18117503720006686,
+  "lr": 8.135394049241399e-05,
+  "weight_decay": 0.015076469897649124,
+  "backbone_lr_ratio": 0.9586175198423679,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_2/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 11, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 32, 33, 34, 36, 37, 38, 39, 40, 41, 43, 44, 45, 47, 48, 49, 50, 51, 55, 56, 57, 58, 59, 60, 61, 63, 64, 65, 68, 69, 70, 72, 74, 76, 78, 79, 80, 82, 83, 84, 85, 86, 87, 88, 91, 93, 94, 95, 96, 97, 100, 101, 102, 103, 104, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 142, 143, 145, 147, 148, 149, 150, 151, 152, 153, 156, 157, 158, 159, 160, 161, 162, 163, 165, 166, 167, 168, 170, 171, 172, 173, 174, 175, 177, 179, 180, 181, 183, 184, 185, 186, 187, 188, 191, 192, 193, 194, 196, 197, 199, 200, 202, 203, 204, 205, 207, 208, 209, 210, 211, 212, 213, 214, 215, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 232, 233, 235, 236, 237, 238, 239, 240, 241, 242, 243, 245, 246, 249, 250, 251, 252, 253, 254, 255, 256, 257, 259, 260, 261, 263, 264, 265, 266, 267, 268, 269, 271, 272, 273, 274, 275, 276, 280, 281, 282, 283, 284, 285, 287, 289, 290, 293, 294, 295, 296, 297, 298, 299, 300, 302, 303, 305, 307, 308, 309, 310, 311, 312, 313, 314, 315, 317, 318, 320, 321, 322, 323, 324, 326, 327, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 358, 360, 361, 363, 364, 365, 366, 368, 369, 370, 371, 372, 373, 376, 377, 378, 379, 380, 381, 382, 383, 385, 386, 387, 388, 389, 390, 392, 394, 395, 398, 399, 400, 401, 402, 404, 405, 407, 408, 409, 410, 412, 413, 414, 415, 416, 417, 418], "outer_test_idx": [10, 12, 15, 31, 35, 42, 46, 52, 53, 54, 62, 66, 67, 71, 73, 75, 77, 81, 89, 90, 92, 98, 99, 105, 123, 124, 141, 144, 146, 154, 155, 164, 169, 176, 178, 182, 189, 190, 195, 198, 201, 206, 216, 217, 218, 219, 231, 234, 244, 247, 248, 258, 262, 270, 277, 278, 279, 286, 288, 291, 292, 301, 304, 306, 316, 319, 325, 328, 343, 357, 359, 362, 367, 374, 375, 384, 391, 393, 396, 397, 403, 406, 411, 419]}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_2/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.10285134147020851,
+    "rmse": 0.3207044456664243,
+    "mae": 0.2692902485529582,
+    "r2": -0.10201936510861498
+  },
+  "delivery": {
+    "n_samples": 60,
+    "mse": 0.663864171054346,
+    "rmse": 0.8147786024769833,
+    "mae": 0.548082401487045,
+    "r2": 0.11628172479876508
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5357142857142857,
+    "precision": 0.4457671957671958,
+    "recall": 0.7063492063492064,
+    "f1": 0.44368858654572946
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6666666666666666,
+    "precision": 0.6064246211305034,
+    "recall": 0.6547619047619048,
+    "f1": 0.6168530247347982
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9344262295081968,
+    "precision": 0.7142857142857143,
+    "recall": 0.9655172413793103,
+    "f1": 0.7821428571428571
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.3895171114701636,
+    "js_divergence": 0.10296173296198505
+  }
+}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_3/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.1799936812783694,
+  "lr": 0.00033865324369584493,
+  "weight_decay": 0.01440133094015265,
+  "backbone_lr_ratio": 0.3278334190266755,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_3/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 4, 5, 7, 8, 9, 10, 11, 12, 14, 15, 16, 17, 20, 21, 22, 23, 24, 25, 27, 28, 29, 30, 31, 32, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 45, 46, 47, 51, 52, 53, 54, 55, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 73, 74, 75, 76, 77, 78, 79, 81, 82, 83, 84, 85, 86, 87, 89, 90, 91, 92, 93, 95, 96, 97, 98, 99, 100, 102, 104, 105, 106, 109, 110, 111, 112, 113, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 127, 129, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 146, 147, 148, 150, 151, 152, 153, 154, 155, 156, 158, 159, 160, 161, 162, 163, 164, 165, 166, 168, 169, 170, 172, 173, 174, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 189, 190, 191, 192, 193, 194, 195, 198, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 212, 213, 214, 215, 216, 217, 218, 219, 221, 223, 224, 226, 227, 228, 230, 231, 233, 234, 235, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 251, 252, 253, 254, 255, 256, 257, 258, 259, 262, 263, 264, 265, 266, 267, 268, 269, 270, 272, 273, 274, 275, 276, 277, 278, 279, 280, 282, 284, 286, 288, 289, 290, 291, 292, 293, 294, 295, 297, 298, 299, 301, 302, 304, 305, 306, 307, 308, 309, 310, 312, 314, 315, 316, 317, 319, 320, 321, 325, 326, 327, 328, 331, 332, 334, 336, 340, 341, 343, 344, 346, 347, 348, 350, 352, 353, 357, 359, 360, 361, 362, 363, 364, 366, 367, 368, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 382, 383, 384, 385, 386, 387, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 400, 401, 402, 403, 404, 406, 407, 410, 411, 412, 414, 415, 416, 418, 419], "outer_test_idx": [3, 6, 13, 18, 19, 26, 33, 44, 48, 49, 50, 56, 57, 72, 80, 88, 94, 101, 103, 107, 108, 114, 126, 128, 130, 145, 149, 157, 167, 171, 175, 188, 196, 197, 199, 211, 220, 222, 225, 229, 232, 236, 250, 260, 261, 271, 281, 283, 285, 287, 296, 300, 303, 311, 313, 318, 322, 323, 324, 329, 330, 333, 335, 337, 338, 339, 342, 345, 349, 351, 354, 355, 356, 358, 365, 369, 381, 388, 399, 405, 408, 409, 413, 417]}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_3/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.4783278330356588,
+    "rmse": 0.6916124876226996,
+    "mae": 0.3700549229081855,
+    "r2": -0.4677665928655337
+  },
+  "delivery": {
+    "n_samples": 59,
+    "mse": 0.8242563169380902,
+    "rmse": 0.907885629877514,
+    "mae": 0.6424747443900017,
+    "r2": 0.19872492929644914
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.7023809523809523,
+    "precision": 0.4810152874709206,
+    "recall": 0.7291346646185355,
+    "f1": 0.48990704493096837
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6428571428571429,
+    "precision": 0.5627865961199294,
+    "recall": 0.5790095525389644,
+    "f1": 0.5641184865618634
+  },
+  "toxic": {
+    "n_samples": 60,
+    "accuracy": 0.9333333333333333,
+    "precision": 0.7142857142857143,
+    "recall": 0.9649122807017544,
+    "f1": 0.7818181818181817
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.40656307520365187,
+    "js_divergence": 0.10906334576015857
+  }
+}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_4/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.14162975892531532,
+  "lr": 0.0004019907813153378,
+  "weight_decay": 0.00037960139235180325,
+  "backbone_lr_ratio": 0.18010931581670617,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_4/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [1, 2, 3, 4, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17, 18, 19, 20, 22, 23, 25, 26, 28, 29, 31, 33, 34, 35, 37, 41, 42, 43, 44, 45, 46, 48, 49, 50, 51, 52, 53, 54, 56, 57, 59, 60, 62, 63, 64, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 77, 78, 79, 80, 81, 82, 83, 84, 85, 87, 88, 89, 90, 91, 92, 93, 94, 96, 97, 98, 99, 101, 102, 103, 104, 105, 107, 108, 110, 111, 112, 113, 114, 116, 117, 119, 120, 121, 122, 123, 124, 126, 127, 128, 130, 132, 133, 134, 135, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 149, 152, 154, 155, 157, 160, 161, 162, 163, 164, 165, 166, 167, 169, 170, 171, 173, 174, 175, 176, 177, 178, 180, 181, 182, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 205, 206, 207, 208, 209, 211, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 234, 235, 236, 238, 239, 240, 241, 243, 244, 245, 246, 247, 248, 249, 250, 251, 253, 254, 257, 258, 260, 261, 262, 263, 264, 265, 266, 267, 270, 271, 272, 273, 277, 278, 279, 280, 281, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 298, 299, 300, 301, 302, 303, 304, 305, 306, 309, 311, 312, 313, 314, 315, 316, 318, 319, 320, 322, 323, 324, 325, 326, 328, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 341, 342, 343, 345, 346, 347, 349, 350, 351, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 365, 367, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 381, 383, 384, 385, 386, 388, 390, 391, 392, 393, 394, 395, 396, 397, 399, 400, 401, 402, 403, 405, 406, 407, 408, 409, 410, 411, 413, 415, 416, 417, 418, 419], "outer_test_idx": [0, 5, 11, 21, 24, 27, 30, 32, 36, 38, 39, 40, 47, 55, 58, 61, 65, 76, 86, 95, 100, 106, 109, 115, 118, 125, 129, 131, 136, 148, 150, 151, 153, 156, 158, 159, 168, 172, 179, 183, 184, 185, 202, 204, 210, 212, 213, 228, 233, 237, 242, 252, 255, 256, 259, 268, 269, 274, 275, 276, 282, 297, 307, 308, 310, 317, 321, 327, 340, 344, 348, 352, 364, 366, 368, 379, 380, 382, 387, 389, 398, 404, 412, 414]}

models/abl/baseline/20260617_180556/repeat_1/outer_fold_4/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.16667444743348794,
+    "rmse": 0.4082578198069058,
+    "mae": 0.2857146007674081,
+    "r2": -0.20998938958412183
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.9768153651071599,
+    "rmse": 0.9883397012703476,
+    "mae": 0.6390594764259355,
+    "r2": 0.055435667309009284
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.6904761904761905,
+    "precision": 0.49700048379293665,
+    "recall": 0.7552483358934973,
+    "f1": 0.533806146572104
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6904761904761905,
+    "precision": 0.6299589603283173,
+    "recall": 0.6871794871794873,
+    "f1": 0.6483638563493374
+  },
+  "toxic": {
+    "n_samples": 59,
+    "accuracy": 0.9661016949152542,
+    "precision": 0.8,
+    "recall": 0.9821428571428572,
+    "f1": 0.8659090909090909
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.394783094910233,
+    "js_divergence": 0.10592993993751994
+  }
+}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_0/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.28242799368681437,
+  "lr": 0.00037183641805732076,
+  "weight_decay": 6.290644294586152e-05,
+  "backbone_lr_ratio": 0.10677482709481352,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_0/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 19, 21, 22, 23, 24, 25, 26, 27, 28, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 52, 53, 54, 55, 56, 57, 58, 61, 62, 63, 64, 65, 66, 67, 68, 71, 72, 73, 75, 76, 77, 80, 81, 85, 86, 88, 89, 90, 92, 94, 95, 96, 98, 99, 100, 101, 103, 104, 105, 106, 107, 108, 109, 111, 113, 114, 115, 116, 117, 118, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 133, 134, 135, 136, 138, 139, 141, 144, 145, 146, 147, 148, 149, 150, 151, 153, 154, 155, 156, 157, 158, 159, 160, 163, 164, 167, 168, 169, 170, 171, 172, 175, 176, 178, 179, 180, 182, 183, 184, 185, 188, 189, 190, 191, 192, 193, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 206, 208, 210, 211, 212, 213, 216, 217, 218, 219, 220, 222, 224, 225, 227, 228, 229, 230, 231, 232, 233, 234, 236, 237, 239, 241, 242, 244, 245, 247, 248, 249, 250, 252, 253, 255, 256, 257, 258, 259, 260, 261, 262, 265, 267, 268, 269, 270, 271, 272, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 285, 286, 287, 288, 291, 292, 294, 296, 297, 298, 299, 300, 301, 303, 304, 306, 307, 308, 309, 310, 311, 312, 313, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329, 330, 332, 333, 335, 336, 337, 338, 339, 340, 342, 343, 344, 345, 346, 347, 348, 349, 351, 352, 354, 355, 356, 357, 358, 359, 360, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 374, 375, 377, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389, 390, 391, 393, 395, 396, 397, 398, 399, 400, 401, 403, 404, 405, 406, 407, 408, 409, 411, 412, 413, 414, 415, 417, 418, 419], "outer_test_idx": [7, 16, 20, 29, 51, 59, 60, 69, 70, 74, 78, 79, 82, 83, 84, 87, 91, 93, 97, 102, 110, 112, 119, 120, 121, 132, 137, 140, 142, 143, 152, 161, 162, 165, 166, 173, 174, 177, 181, 186, 187, 194, 205, 207, 209, 214, 215, 221, 223, 226, 235, 238, 240, 243, 246, 251, 254, 263, 264, 266, 273, 284, 289, 290, 293, 295, 302, 305, 314, 315, 331, 334, 341, 350, 353, 361, 373, 376, 378, 392, 394, 402, 410, 416]}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_0/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.38432886654317966,
+    "rmse": 0.6199426316548812,
+    "mae": 0.43342806632260245,
+    "r2": -0.2761097384806048
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.6744019580730524,
+    "rmse": 0.8212197988803317,
+    "mae": 0.6701494166306381,
+    "r2": 0.14279441523317304
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5357142857142857,
+    "precision": 0.35634973404255316,
+    "recall": 0.547423887587822,
+    "f1": 0.3495414294750159
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6904761904761905,
+    "precision": 0.6587301587301587,
+    "recall": 0.6938095238095238,
+    "f1": 0.6526081185986528
+  },
+  "toxic": {
+    "n_samples": 58,
+    "accuracy": 0.9655172413793104,
+    "precision": 0.75,
+    "recall": 0.9821428571428572,
+    "f1": 0.8242424242424242
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.40697274862376626,
+    "js_divergence": 0.09501893155280818
+  }
+}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_1/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.23085562232445592,
+  "lr": 0.0005227270589225132,
+  "weight_decay": 0.00461701039547356,
+  "backbone_lr_ratio": 0.12087164274488635,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_1/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 3, 5, 6, 7, 10, 11, 12, 13, 15, 16, 18, 19, 20, 21, 24, 26, 27, 29, 30, 31, 32, 33, 35, 36, 38, 39, 40, 42, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 65, 66, 67, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 97, 98, 99, 100, 101, 102, 103, 105, 106, 107, 108, 109, 110, 112, 114, 115, 118, 119, 120, 121, 123, 124, 125, 126, 128, 129, 130, 131, 132, 136, 137, 140, 141, 142, 143, 144, 145, 146, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 161, 162, 164, 165, 166, 167, 168, 169, 171, 172, 173, 174, 175, 176, 177, 178, 179, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 194, 195, 196, 197, 198, 199, 201, 202, 204, 205, 206, 207, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 225, 226, 228, 229, 231, 232, 233, 234, 235, 236, 237, 238, 240, 242, 243, 244, 246, 247, 248, 250, 251, 252, 254, 255, 256, 258, 259, 260, 261, 262, 263, 264, 266, 268, 269, 270, 271, 273, 274, 275, 276, 277, 278, 279, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 295, 296, 297, 300, 301, 302, 303, 304, 305, 306, 307, 308, 310, 311, 313, 314, 315, 316, 317, 318, 319, 321, 322, 323, 324, 325, 327, 328, 329, 330, 331, 333, 334, 335, 337, 338, 339, 340, 341, 342, 343, 344, 345, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 361, 362, 364, 365, 366, 367, 368, 369, 373, 374, 375, 376, 378, 379, 380, 381, 382, 384, 387, 388, 389, 391, 392, 393, 394, 396, 397, 398, 399, 402, 403, 404, 405, 406, 408, 409, 410, 411, 412, 413, 414, 416, 417, 419], "outer_test_idx": [1, 2, 4, 8, 9, 14, 17, 22, 23, 25, 28, 34, 37, 41, 43, 45, 63, 64, 68, 85, 96, 104, 111, 113, 116, 117, 122, 127, 133, 134, 135, 138, 139, 147, 160, 163, 170, 180, 191, 192, 193, 200, 203, 208, 224, 227, 230, 239, 241, 245, 249, 253, 257, 265, 267, 272, 280, 294, 298, 299, 309, 312, 320, 326, 332, 336, 346, 347, 360, 363, 370, 371, 372, 377, 383, 385, 386, 390, 395, 400, 401, 407, 415, 418]}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_1/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.28709899653313253,
+    "rmse": 0.5358161965946275,
+    "mae": 0.41961162147067843,
+    "r2": -3.7562675545651993
+  },
+  "delivery": {
+    "n_samples": 61,
+    "mse": 1.0055884332484593,
+    "rmse": 1.0027903236711349,
+    "mae": 0.6592718309371687,
+    "r2": 0.23884204782685936
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.44047619047619047,
+    "precision": 0.34375,
+    "recall": 0.32275132275132273,
+    "f1": 0.2913806254767353
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.5357142857142857,
+    "precision": 0.5,
+    "recall": 0.5004995004995004,
+    "f1": 0.4796109240553685
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9508196721311475,
+    "precision": 0.75,
+    "recall": 0.9741379310344828,
+    "f1": 0.8200589970501475
+  },
+  "biodist": {
+    "n_samples": 61,
+    "kl_divergence": 0.4974494276121329,
+    "js_divergence": 0.1259321573319038
+  }
+}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_2/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.18117503720006686,
+  "lr": 8.135394049241399e-05,
+  "weight_decay": 0.015076469897649124,
+  "backbone_lr_ratio": 0.9586175198423679,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_2/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 11, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 32, 33, 34, 36, 37, 38, 39, 40, 41, 43, 44, 45, 47, 48, 49, 50, 51, 55, 56, 57, 58, 59, 60, 61, 63, 64, 65, 68, 69, 70, 72, 74, 76, 78, 79, 80, 82, 83, 84, 85, 86, 87, 88, 91, 93, 94, 95, 96, 97, 100, 101, 102, 103, 104, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 142, 143, 145, 147, 148, 149, 150, 151, 152, 153, 156, 157, 158, 159, 160, 161, 162, 163, 165, 166, 167, 168, 170, 171, 172, 173, 174, 175, 177, 179, 180, 181, 183, 184, 185, 186, 187, 188, 191, 192, 193, 194, 196, 197, 199, 200, 202, 203, 204, 205, 207, 208, 209, 210, 211, 212, 213, 214, 215, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 232, 233, 235, 236, 237, 238, 239, 240, 241, 242, 243, 245, 246, 249, 250, 251, 252, 253, 254, 255, 256, 257, 259, 260, 261, 263, 264, 265, 266, 267, 268, 269, 271, 272, 273, 274, 275, 276, 280, 281, 282, 283, 284, 285, 287, 289, 290, 293, 294, 295, 296, 297, 298, 299, 300, 302, 303, 305, 307, 308, 309, 310, 311, 312, 313, 314, 315, 317, 318, 320, 321, 322, 323, 324, 326, 327, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 358, 360, 361, 363, 364, 365, 366, 368, 369, 370, 371, 372, 373, 376, 377, 378, 379, 380, 381, 382, 383, 385, 386, 387, 388, 389, 390, 392, 394, 395, 398, 399, 400, 401, 402, 404, 405, 407, 408, 409, 410, 412, 413, 414, 415, 416, 417, 418], "outer_test_idx": [10, 12, 15, 31, 35, 42, 46, 52, 53, 54, 62, 66, 67, 71, 73, 75, 77, 81, 89, 90, 92, 98, 99, 105, 123, 124, 141, 144, 146, 154, 155, 164, 169, 176, 178, 182, 189, 190, 195, 198, 201, 206, 216, 217, 218, 219, 231, 234, 244, 247, 248, 258, 262, 270, 277, 278, 279, 286, 288, 291, 292, 301, 304, 306, 316, 319, 325, 328, 343, 357, 359, 362, 367, 374, 375, 384, 391, 393, 396, 397, 403, 406, 411, 419]}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_2/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.12734599676966307,
+    "rmse": 0.3568557086129674,
+    "mae": 0.283481532619113,
+    "r2": -0.3644717949534717
+  },
+  "delivery": {
+    "n_samples": 60,
+    "mse": 0.6995736506531275,
+    "rmse": 0.8364051952571359,
+    "mae": 0.5743296845893686,
+    "r2": 0.06874621814647808
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5119047619047619,
+    "precision": 0.41008771929824556,
+    "recall": 0.6746031746031745,
+    "f1": 0.4050345260514752
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6666666666666666,
+    "precision": 0.6189974937343359,
+    "recall": 0.687908496732026,
+    "f1": 0.6319077437787746
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9344262295081968,
+    "precision": 0.7142857142857143,
+    "recall": 0.9655172413793103,
+    "f1": 0.7821428571428571
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.3964743242871827,
+    "js_divergence": 0.10757710661599443
+  }
+}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_3/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.1799936812783694,
+  "lr": 0.00033865324369584493,
+  "weight_decay": 0.01440133094015265,
+  "backbone_lr_ratio": 0.3278334190266755,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_3/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 4, 5, 7, 8, 9, 10, 11, 12, 14, 15, 16, 17, 20, 21, 22, 23, 24, 25, 27, 28, 29, 30, 31, 32, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 45, 46, 47, 51, 52, 53, 54, 55, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 73, 74, 75, 76, 77, 78, 79, 81, 82, 83, 84, 85, 86, 87, 89, 90, 91, 92, 93, 95, 96, 97, 98, 99, 100, 102, 104, 105, 106, 109, 110, 111, 112, 113, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 127, 129, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 146, 147, 148, 150, 151, 152, 153, 154, 155, 156, 158, 159, 160, 161, 162, 163, 164, 165, 166, 168, 169, 170, 172, 173, 174, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 189, 190, 191, 192, 193, 194, 195, 198, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 212, 213, 214, 215, 216, 217, 218, 219, 221, 223, 224, 226, 227, 228, 230, 231, 233, 234, 235, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 251, 252, 253, 254, 255, 256, 257, 258, 259, 262, 263, 264, 265, 266, 267, 268, 269, 270, 272, 273, 274, 275, 276, 277, 278, 279, 280, 282, 284, 286, 288, 289, 290, 291, 292, 293, 294, 295, 297, 298, 299, 301, 302, 304, 305, 306, 307, 308, 309, 310, 312, 314, 315, 316, 317, 319, 320, 321, 325, 326, 327, 328, 331, 332, 334, 336, 340, 341, 343, 344, 346, 347, 348, 350, 352, 353, 357, 359, 360, 361, 362, 363, 364, 366, 367, 368, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 382, 383, 384, 385, 386, 387, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 400, 401, 402, 403, 404, 406, 407, 410, 411, 412, 414, 415, 416, 418, 419], "outer_test_idx": [3, 6, 13, 18, 19, 26, 33, 44, 48, 49, 50, 56, 57, 72, 80, 88, 94, 101, 103, 107, 108, 114, 126, 128, 130, 145, 149, 157, 167, 171, 175, 188, 196, 197, 199, 211, 220, 222, 225, 229, 232, 236, 250, 260, 261, 271, 281, 283, 285, 287, 296, 300, 303, 311, 313, 318, 322, 323, 324, 329, 330, 333, 335, 337, 338, 339, 342, 345, 349, 351, 354, 355, 356, 358, 365, 369, 381, 388, 399, 405, 408, 409, 413, 417]}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_3/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.46358638894813786,
+    "rmse": 0.6808717859833361,
+    "mae": 0.37019611553973464,
+    "r2": -0.422531928127458
+  },
+  "delivery": {
+    "n_samples": 59,
+    "mse": 0.8370140113507513,
+    "rmse": 0.9148846983914155,
+    "mae": 0.6370043654657774,
+    "r2": 0.18632293457411198
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.6071428571428571,
+    "precision": 0.47548510048510045,
+    "recall": 0.7491039426523297,
+    "f1": 0.4726173766801967
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6785714285714286,
+    "precision": 0.6115219070742745,
+    "recall": 0.6469582704876823,
+    "f1": 0.6198067632850242
+  },
+  "toxic": {
+    "n_samples": 60,
+    "accuracy": 0.9333333333333333,
+    "precision": 0.7142857142857143,
+    "recall": 0.9649122807017544,
+    "f1": 0.7818181818181817
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.40801997248379873,
+    "js_divergence": 0.11086638268723202
+  }
+}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_4/best_params.json

@@ -0,0 +1,12 @@
+{
+  "dropout": 0.14162975892531532,
+  "lr": 0.0004019907813153378,
+  "weight_decay": 0.00037960139235180325,
+  "backbone_lr_ratio": 0.18010931581670617,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_4/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [1, 2, 3, 4, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17, 18, 19, 20, 22, 23, 25, 26, 28, 29, 31, 33, 34, 35, 37, 41, 42, 43, 44, 45, 46, 48, 49, 50, 51, 52, 53, 54, 56, 57, 59, 60, 62, 63, 64, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 77, 78, 79, 80, 81, 82, 83, 84, 85, 87, 88, 89, 90, 91, 92, 93, 94, 96, 97, 98, 99, 101, 102, 103, 104, 105, 107, 108, 110, 111, 112, 113, 114, 116, 117, 119, 120, 121, 122, 123, 124, 126, 127, 128, 130, 132, 133, 134, 135, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 149, 152, 154, 155, 157, 160, 161, 162, 163, 164, 165, 166, 167, 169, 170, 171, 173, 174, 175, 176, 177, 178, 180, 181, 182, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 205, 206, 207, 208, 209, 211, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 234, 235, 236, 238, 239, 240, 241, 243, 244, 245, 246, 247, 248, 249, 250, 251, 253, 254, 257, 258, 260, 261, 262, 263, 264, 265, 266, 267, 270, 271, 272, 273, 277, 278, 279, 280, 281, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 298, 299, 300, 301, 302, 303, 304, 305, 306, 309, 311, 312, 313, 314, 315, 316, 318, 319, 320, 322, 323, 324, 325, 326, 328, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 341, 342, 343, 345, 346, 347, 349, 350, 351, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 365, 367, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 381, 383, 384, 385, 386, 388, 390, 391, 392, 393, 394, 395, 396, 397, 399, 400, 401, 402, 403, 405, 406, 407, 408, 409, 410, 411, 413, 415, 416, 417, 418, 419], "outer_test_idx": [0, 5, 11, 21, 24, 27, 30, 32, 36, 38, 39, 40, 47, 55, 58, 61, 65, 76, 86, 95, 100, 106, 109, 115, 118, 125, 129, 131, 136, 148, 150, 151, 153, 156, 158, 159, 168, 172, 179, 183, 184, 185, 202, 204, 210, 212, 213, 228, 233, 237, 242, 252, 255, 256, 259, 268, 269, 274, 275, 276, 282, 297, 307, 308, 310, 317, 321, 327, 340, 344, 348, 352, 364, 366, 368, 379, 380, 382, 387, 389, 398, 404, 412, 414]}

models/abl/baseline/20260617_180556/repeat_2/outer_fold_4/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.15145313313910996,
+    "rmse": 0.3891697998806048,
+    "mae": 0.2744193644750686,
+    "r2": -0.09948877551085711
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.9378577638427679,
+    "rmse": 0.9684305673835207,
+    "mae": 0.6135335185661398,
+    "r2": 0.09310702461562304
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.6071428571428571,
+    "precision": 0.4863247863247863,
+    "recall": 0.7491039426523297,
+    "f1": 0.5019607843137255
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6785714285714286,
+    "precision": 0.6324786324786325,
+    "recall": 0.7021367521367522,
+    "f1": 0.6488452445899254
+  },
+  "toxic": {
+    "n_samples": 59,
+    "accuracy": 0.9661016949152542,
+    "precision": 0.8,
+    "recall": 0.9821428571428572,
+    "f1": 0.8659090909090909
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.39116454524118993,
+    "js_divergence": 0.1065009133663494
+  }
+}

models/abl/baseline/20260617_180556/strata_info.json

@@ -0,0 +1,60 @@
+{
+  "original_strata_counts": {
+    "T0|P0|E0": "5",
+    "T0|P0|E1": "58",
+    "T0|P0|E2": "169",
+    "T0|P1|E0": "1",
+    "T0|P1|E1": "17",
+    "T0|P1|E2": "32",
+    "T0|P2|E2": "3",
+    "T1|P0|E2": "9",
+    "T1|P1|E2": "5",
+    "TNA|P0|E0": "28",
+    "TNA|P0|E1": "21",
+    "TNA|P0|E2": "20",
+    "TNA|P1|E0": "29",
+    "TNA|P1|E1": "7",
+    "TNA|P1|E2": "14",
+    "TNA|P2|E2": "1",
+    "TNA|P3|E0": "1"
+  },
+  "rare_strata": [
+    "T0|P1|E0",
+    "T0|P2|E2",
+    "TNA|P2|E2",
+    "TNA|P3|E0"
+  ],
+  "final_strata": [
+    "RARE",
+    "T0|P0|E0",
+    "T0|P0|E1",
+    "T0|P0|E2",
+    "T0|P1|E1",
+    "T0|P1|E2",
+    "T1|P0|E2",
+    "T1|P1|E2",
+    "TNA|P0|E0",
+    "TNA|P0|E1",
+    "TNA|P0|E2",
+    "TNA|P1|E0",
+    "TNA|P1|E1",
+    "TNA|P1|E2"
+  ],
+  "final_strata_counts": {
+    "RARE": "6",
+    "T0|P0|E0": "5",
+    "T0|P0|E1": "58",
+    "T0|P0|E2": "169",
+    "T0|P1|E1": "17",
+    "T0|P1|E2": "32",
+    "T1|P0|E2": "9",
+    "T1|P1|E2": "5",
+    "TNA|P0|E0": "28",
+    "TNA|P0|E1": "21",
+    "TNA|P0|E2": "20",
+    "TNA|P1|E0": "29",
+    "TNA|P1|E1": "7",
+    "TNA|P1|E2": "14"
+  },
+  "n_rare_merged": "6"
+}

models/abl/baseline/20260617_180556/summary.json

@@ -0,0 +1,932 @@
+{
+  "fold_results": [
+    {
+      "fold": 0,
+      "best_params": {
+        "dropout": 0.28242799368681437,
+        "lr": 0.00037183641805732076,
+        "weight_decay": 6.290644294586152e-05,
+        "backbone_lr_ratio": 0.10677482709481352,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 15,
+      "test_metrics": {
+        "size": {
+          "n_samples": 83,
+          "mse": 0.32790454280151365,
+          "rmse": 0.572629498717551,
+          "mae": 0.35871042688208893,
+          "r2": -0.08876073797082196
+        },
+        "delivery": {
+          "n_samples": 58,
+          "mse": 0.7250728333301815,
+          "rmse": 0.8515120864263651,
+          "mae": 0.7098816664696768,
+          "r2": 0.07838867510225656
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.5476190476190477,
+          "precision": 0.3435971685971686,
+          "recall": 0.5359094457455114,
+          "f1": 0.34977324263038545
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.7142857142857143,
+          "precision": 0.6871101871101871,
+          "recall": 0.707142857142857,
+          "f1": 0.6756066411238826
+        },
+        "toxic": {
+          "n_samples": 58,
+          "accuracy": 0.9655172413793104,
+          "precision": 0.75,
+          "recall": 0.9821428571428572,
+          "f1": 0.8242424242424242
+        },
+        "biodist": {
+          "n_samples": 58,
+          "kl_divergence": 0.4073394583325288,
+          "js_divergence": 0.092887269703335
+        }
+      }
+    },
+    {
+      "fold": 1,
+      "best_params": {
+        "dropout": 0.23085562232445592,
+        "lr": 0.0005227270589225132,
+        "weight_decay": 0.00461701039547356,
+        "backbone_lr_ratio": 0.12087164274488635,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 9,
+      "test_metrics": {
+        "size": {
+          "n_samples": 84,
+          "mse": 0.233876145855699,
+          "rmse": 0.4836074294876982,
+          "mae": 0.3760785318556286,
+          "r2": -2.874543407510119
+        },
+        "delivery": {
+          "n_samples": 61,
+          "mse": 1.0238964615450643,
+          "rmse": 1.0118776910007772,
+          "mae": 0.6578027546466862,
+          "r2": 0.2249841902125319
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.5238095238095238,
+          "precision": 0.34497354497354493,
+          "recall": 0.3280423280423281,
+          "f1": 0.3231922398589065
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.5595238095238095,
+          "precision": 0.5101731601731602,
+          "recall": 0.5036154321868608,
+          "f1": 0.48790324563862697
+        },
+        "toxic": {
+          "n_samples": 61,
+          "accuracy": 0.9508196721311475,
+          "precision": 0.75,
+          "recall": 0.9741379310344828,
+          "f1": 0.8200589970501475
+        },
+        "biodist": {
+          "n_samples": 61,
+          "kl_divergence": 0.48528378254405913,
+          "js_divergence": 0.12258377401744426
+        }
+      }
+    },
+    {
+      "fold": 2,
+      "best_params": {
+        "dropout": 0.18117503720006686,
+        "lr": 8.135394049241399e-05,
+        "weight_decay": 0.015076469897649124,
+        "backbone_lr_ratio": 0.9586175198423679,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 22,
+      "test_metrics": {
+        "size": {
+          "n_samples": 84,
+          "mse": 0.07870736214858826,
+          "rmse": 0.28054832408800495,
+          "mae": 0.2252055633635748,
+          "r2": 0.15667568332606574
+        },
+        "delivery": {
+          "n_samples": 60,
+          "mse": 0.6624288051399898,
+          "rmse": 0.8138972939750997,
+          "mae": 0.5578871982870623,
+          "r2": 0.118192445613993
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.5714285714285714,
+          "precision": 0.43453195231581004,
+          "recall": 0.7014848950332823,
+          "f1": 0.4452380952380952
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.6666666666666666,
+          "precision": 0.6140109890109889,
+          "recall": 0.6785714285714285,
+          "f1": 0.6262626262626263
+        },
+        "toxic": {
+          "n_samples": 61,
+          "accuracy": 0.9344262295081968,
+          "precision": 0.7142857142857143,
+          "recall": 0.9655172413793103,
+          "f1": 0.7821428571428571
+        },
+        "biodist": {
+          "n_samples": 60,
+          "kl_divergence": 0.3967825471495568,
+          "js_divergence": 0.1057568924841401
+        }
+      }
+    },
+    {
+      "fold": 3,
+      "best_params": {
+        "dropout": 0.1799936812783694,
+        "lr": 0.00033865324369584493,
+        "weight_decay": 0.01440133094015265,
+        "backbone_lr_ratio": 0.3278334190266755,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 10,
+      "test_metrics": {
+        "size": {
+          "n_samples": 83,
+          "mse": 0.477705869562307,
+          "rmse": 0.6911626939891266,
+          "mae": 0.38659381004701177,
+          "r2": -0.46585807501413723
+        },
+        "delivery": {
+          "n_samples": 59,
+          "mse": 0.7692103143544243,
+          "rmse": 0.8770463581558413,
+          "mae": 0.6155587352528158,
+          "r2": 0.2522361838732067
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.6904761904761905,
+          "precision": 0.4426767676767677,
+          "recall": 0.6817716333845366,
+          "f1": 0.43613724611708476
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.6428571428571429,
+          "precision": 0.5650987224157956,
+          "recall": 0.5790095525389644,
+          "f1": 0.5637508626639062
+        },
+        "toxic": {
+          "n_samples": 60,
+          "accuracy": 0.9333333333333333,
+          "precision": 0.7142857142857143,
+          "recall": 0.9649122807017544,
+          "f1": 0.7818181818181817
+        },
+        "biodist": {
+          "n_samples": 60,
+          "kl_divergence": 0.4102080974965646,
+          "js_divergence": 0.11100613154873733
+        }
+      }
+    },
+    {
+      "fold": 4,
+      "best_params": {
+        "dropout": 0.14162975892531532,
+        "lr": 0.0004019907813153378,
+        "weight_decay": 0.00037960139235180325,
+        "backbone_lr_ratio": 0.18010931581670617,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 12,
+      "test_metrics": {
+        "size": {
+          "n_samples": 84,
+          "mse": 0.1842198466764517,
+          "rmse": 0.4292083953937198,
+          "mae": 0.29178874833243235,
+          "r2": -0.33736192476816806
+        },
+        "delivery": {
+          "n_samples": 58,
+          "mse": 0.9003281069383875,
+          "rmse": 0.9488562098328637,
+          "mae": 0.6070586243837044,
+          "r2": 0.1293975832987566
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.5476190476190477,
+          "precision": 0.34548450889914306,
+          "recall": 0.5337941628264209,
+          "f1": 0.3497882046269143
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.7023809523809523,
+          "precision": 0.6548765244417418,
+          "recall": 0.7576923076923077,
+          "f1": 0.6777508856852221
+        },
+        "toxic": {
+          "n_samples": 59,
+          "accuracy": 0.9661016949152542,
+          "precision": 0.8,
+          "recall": 0.9821428571428572,
+          "f1": 0.8659090909090909
+        },
+        "biodist": {
+          "n_samples": 58,
+          "kl_divergence": 0.3946735882065513,
+          "js_divergence": 0.1067245125767987
+        }
+      }
+    },
+    {
+      "fold": 0,
+      "best_params": {
+        "dropout": 0.28242799368681437,
+        "lr": 0.00037183641805732076,
+        "weight_decay": 6.290644294586152e-05,
+        "backbone_lr_ratio": 0.10677482709481352,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 15,
+      "test_metrics": {
+        "size": {
+          "n_samples": 83,
+          "mse": 0.3808128541133737,
+          "rmse": 0.6171003598389598,
+          "mae": 0.4168719153806388,
+          "r2": -0.2644353156285024
+        },
+        "delivery": {
+          "n_samples": 58,
+          "mse": 0.751610048614176,
+          "rmse": 0.8669544674400012,
+          "mae": 0.7122741444999802,
+          "r2": 0.04465827311677528
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.5,
+          "precision": 0.3181788931788932,
+          "recall": 0.5039032006245121,
+          "f1": 0.32369614512471656
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.6666666666666666,
+          "precision": 0.6258687258687259,
+          "recall": 0.6461904761904762,
+          "f1": 0.6127803231251508
+        },
+        "toxic": {
+          "n_samples": 58,
+          "accuracy": 0.9655172413793104,
+          "precision": 0.75,
+          "recall": 0.9821428571428572,
+          "f1": 0.8242424242424242
+        },
+        "biodist": {
+          "n_samples": 58,
+          "kl_divergence": 0.39644445917049376,
+          "js_divergence": 0.09334798400608228
+        }
+      }
+    },
+    {
+      "fold": 1,
+      "best_params": {
+        "dropout": 0.23085562232445592,
+        "lr": 0.0005227270589225132,
+        "weight_decay": 0.00461701039547356,
+        "backbone_lr_ratio": 0.12087164274488635,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 9,
+      "test_metrics": {
+        "size": {
+          "n_samples": 84,
+          "mse": 0.3074589292084139,
+          "rmse": 0.5544897917982025,
+          "mae": 0.3793355481965201,
+          "r2": -4.093563359726248
+        },
+        "delivery": {
+          "n_samples": 61,
+          "mse": 1.0592195690639976,
+          "rmse": 1.029183933543464,
+          "mae": 0.6786946221025753,
+          "r2": 0.19824714422578682
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.5952380952380952,
+          "precision": 0.37142213207121927,
+          "recall": 0.35978835978835977,
+          "f1": 0.35976608187134507
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.5476190476190477,
+          "precision": 0.5052287581699346,
+          "recall": 0.5094429380143666,
+          "f1": 0.4903811048389362
+        },
+        "toxic": {
+          "n_samples": 61,
+          "accuracy": 0.9508196721311475,
+          "precision": 0.75,
+          "recall": 0.9741379310344828,
+          "f1": 0.8200589970501475
+        },
+        "biodist": {
+          "n_samples": 61,
+          "kl_divergence": 0.5219711265981701,
+          "js_divergence": 0.13035793372268764
+        }
+      }
+    },
+    {
+      "fold": 2,
+      "best_params": {
+        "dropout": 0.18117503720006686,
+        "lr": 8.135394049241399e-05,
+        "weight_decay": 0.015076469897649124,
+        "backbone_lr_ratio": 0.9586175198423679,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 22,
+      "test_metrics": {
+        "size": {
+          "n_samples": 84,
+          "mse": 0.10285134147020851,
+          "rmse": 0.3207044456664243,
+          "mae": 0.2692902485529582,
+          "r2": -0.10201936510861498
+        },
+        "delivery": {
+          "n_samples": 60,
+          "mse": 0.663864171054346,
+          "rmse": 0.8147786024769833,
+          "mae": 0.548082401487045,
+          "r2": 0.11628172479876508
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.5357142857142857,
+          "precision": 0.4457671957671958,
+          "recall": 0.7063492063492064,
+          "f1": 0.44368858654572946
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.6666666666666666,
+          "precision": 0.6064246211305034,
+          "recall": 0.6547619047619048,
+          "f1": 0.6168530247347982
+        },
+        "toxic": {
+          "n_samples": 61,
+          "accuracy": 0.9344262295081968,
+          "precision": 0.7142857142857143,
+          "recall": 0.9655172413793103,
+          "f1": 0.7821428571428571
+        },
+        "biodist": {
+          "n_samples": 60,
+          "kl_divergence": 0.3895171114701636,
+          "js_divergence": 0.10296173296198505
+        }
+      }
+    },
+    {
+      "fold": 3,
+      "best_params": {
+        "dropout": 0.1799936812783694,
+        "lr": 0.00033865324369584493,
+        "weight_decay": 0.01440133094015265,
+        "backbone_lr_ratio": 0.3278334190266755,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 10,
+      "test_metrics": {
+        "size": {
+          "n_samples": 83,
+          "mse": 0.4783278330356588,
+          "rmse": 0.6916124876226996,
+          "mae": 0.3700549229081855,
+          "r2": -0.4677665928655337
+        },
+        "delivery": {
+          "n_samples": 59,
+          "mse": 0.8242563169380902,
+          "rmse": 0.907885629877514,
+          "mae": 0.6424747443900017,
+          "r2": 0.19872492929644914
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.7023809523809523,
+          "precision": 0.4810152874709206,
+          "recall": 0.7291346646185355,
+          "f1": 0.48990704493096837
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.6428571428571429,
+          "precision": 0.5627865961199294,
+          "recall": 0.5790095525389644,
+          "f1": 0.5641184865618634
+        },
+        "toxic": {
+          "n_samples": 60,
+          "accuracy": 0.9333333333333333,
+          "precision": 0.7142857142857143,
+          "recall": 0.9649122807017544,
+          "f1": 0.7818181818181817
+        },
+        "biodist": {
+          "n_samples": 60,
+          "kl_divergence": 0.40656307520365187,
+          "js_divergence": 0.10906334576015857
+        }
+      }
+    },
+    {
+      "fold": 4,
+      "best_params": {
+        "dropout": 0.14162975892531532,
+        "lr": 0.0004019907813153378,
+        "weight_decay": 0.00037960139235180325,
+        "backbone_lr_ratio": 0.18010931581670617,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 12,
+      "test_metrics": {
+        "size": {
+          "n_samples": 84,
+          "mse": 0.16667444743348794,
+          "rmse": 0.4082578198069058,
+          "mae": 0.2857146007674081,
+          "r2": -0.20998938958412183
+        },
+        "delivery": {
+          "n_samples": 58,
+          "mse": 0.9768153651071599,
+          "rmse": 0.9883397012703476,
+          "mae": 0.6390594764259355,
+          "r2": 0.055435667309009284
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.6904761904761905,
+          "precision": 0.49700048379293665,
+          "recall": 0.7552483358934973,
+          "f1": 0.533806146572104
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.6904761904761905,
+          "precision": 0.6299589603283173,
+          "recall": 0.6871794871794873,
+          "f1": 0.6483638563493374
+        },
+        "toxic": {
+          "n_samples": 59,
+          "accuracy": 0.9661016949152542,
+          "precision": 0.8,
+          "recall": 0.9821428571428572,
+          "f1": 0.8659090909090909
+        },
+        "biodist": {
+          "n_samples": 58,
+          "kl_divergence": 0.394783094910233,
+          "js_divergence": 0.10592993993751994
+        }
+      }
+    },
+    {
+      "fold": 0,
+      "best_params": {
+        "dropout": 0.28242799368681437,
+        "lr": 0.00037183641805732076,
+        "weight_decay": 6.290644294586152e-05,
+        "backbone_lr_ratio": 0.10677482709481352,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 15,
+      "test_metrics": {
+        "size": {
+          "n_samples": 83,
+          "mse": 0.38432886654317966,
+          "rmse": 0.6199426316548812,
+          "mae": 0.43342806632260245,
+          "r2": -0.2761097384806048
+        },
+        "delivery": {
+          "n_samples": 58,
+          "mse": 0.6744019580730524,
+          "rmse": 0.8212197988803317,
+          "mae": 0.6701494166306381,
+          "r2": 0.14279441523317304
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.5357142857142857,
+          "precision": 0.35634973404255316,
+          "recall": 0.547423887587822,
+          "f1": 0.3495414294750159
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.6904761904761905,
+          "precision": 0.6587301587301587,
+          "recall": 0.6938095238095238,
+          "f1": 0.6526081185986528
+        },
+        "toxic": {
+          "n_samples": 58,
+          "accuracy": 0.9655172413793104,
+          "precision": 0.75,
+          "recall": 0.9821428571428572,
+          "f1": 0.8242424242424242
+        },
+        "biodist": {
+          "n_samples": 58,
+          "kl_divergence": 0.40697274862376626,
+          "js_divergence": 0.09501893155280818
+        }
+      }
+    },
+    {
+      "fold": 1,
+      "best_params": {
+        "dropout": 0.23085562232445592,
+        "lr": 0.0005227270589225132,
+        "weight_decay": 0.00461701039547356,
+        "backbone_lr_ratio": 0.12087164274488635,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 9,
+      "test_metrics": {
+        "size": {
+          "n_samples": 84,
+          "mse": 0.28709899653313253,
+          "rmse": 0.5358161965946275,
+          "mae": 0.41961162147067843,
+          "r2": -3.7562675545651993
+        },
+        "delivery": {
+          "n_samples": 61,
+          "mse": 1.0055884332484593,
+          "rmse": 1.0027903236711349,
+          "mae": 0.6592718309371687,
+          "r2": 0.23884204782685936
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.44047619047619047,
+          "precision": 0.34375,
+          "recall": 0.32275132275132273,
+          "f1": 0.2913806254767353
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.5357142857142857,
+          "precision": 0.5,
+          "recall": 0.5004995004995004,
+          "f1": 0.4796109240553685
+        },
+        "toxic": {
+          "n_samples": 61,
+          "accuracy": 0.9508196721311475,
+          "precision": 0.75,
+          "recall": 0.9741379310344828,
+          "f1": 0.8200589970501475
+        },
+        "biodist": {
+          "n_samples": 61,
+          "kl_divergence": 0.4974494276121329,
+          "js_divergence": 0.1259321573319038
+        }
+      }
+    },
+    {
+      "fold": 2,
+      "best_params": {
+        "dropout": 0.18117503720006686,
+        "lr": 8.135394049241399e-05,
+        "weight_decay": 0.015076469897649124,
+        "backbone_lr_ratio": 0.9586175198423679,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 22,
+      "test_metrics": {
+        "size": {
+          "n_samples": 84,
+          "mse": 0.12734599676966307,
+          "rmse": 0.3568557086129674,
+          "mae": 0.283481532619113,
+          "r2": -0.3644717949534717
+        },
+        "delivery": {
+          "n_samples": 60,
+          "mse": 0.6995736506531275,
+          "rmse": 0.8364051952571359,
+          "mae": 0.5743296845893686,
+          "r2": 0.06874621814647808
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.5119047619047619,
+          "precision": 0.41008771929824556,
+          "recall": 0.6746031746031745,
+          "f1": 0.4050345260514752
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.6666666666666666,
+          "precision": 0.6189974937343359,
+          "recall": 0.687908496732026,
+          "f1": 0.6319077437787746
+        },
+        "toxic": {
+          "n_samples": 61,
+          "accuracy": 0.9344262295081968,
+          "precision": 0.7142857142857143,
+          "recall": 0.9655172413793103,
+          "f1": 0.7821428571428571
+        },
+        "biodist": {
+          "n_samples": 60,
+          "kl_divergence": 0.3964743242871827,
+          "js_divergence": 0.10757710661599443
+        }
+      }
+    },
+    {
+      "fold": 3,
+      "best_params": {
+        "dropout": 0.1799936812783694,
+        "lr": 0.00033865324369584493,
+        "weight_decay": 0.01440133094015265,
+        "backbone_lr_ratio": 0.3278334190266755,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 10,
+      "test_metrics": {
+        "size": {
+          "n_samples": 83,
+          "mse": 0.46358638894813786,
+          "rmse": 0.6808717859833361,
+          "mae": 0.37019611553973464,
+          "r2": -0.422531928127458
+        },
+        "delivery": {
+          "n_samples": 59,
+          "mse": 0.8370140113507513,
+          "rmse": 0.9148846983914155,
+          "mae": 0.6370043654657774,
+          "r2": 0.18632293457411198
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.6071428571428571,
+          "precision": 0.47548510048510045,
+          "recall": 0.7491039426523297,
+          "f1": 0.4726173766801967
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.6785714285714286,
+          "precision": 0.6115219070742745,
+          "recall": 0.6469582704876823,
+          "f1": 0.6198067632850242
+        },
+        "toxic": {
+          "n_samples": 60,
+          "accuracy": 0.9333333333333333,
+          "precision": 0.7142857142857143,
+          "recall": 0.9649122807017544,
+          "f1": 0.7818181818181817
+        },
+        "biodist": {
+          "n_samples": 60,
+          "kl_divergence": 0.40801997248379873,
+          "js_divergence": 0.11086638268723202
+        }
+      }
+    },
+    {
+      "fold": 4,
+      "best_params": {
+        "dropout": 0.14162975892531532,
+        "lr": 0.0004019907813153378,
+        "weight_decay": 0.00037960139235180325,
+        "backbone_lr_ratio": 0.18010931581670617,
+        "d_model": 256,
+        "num_heads": 8,
+        "n_attn_layers": 4,
+        "fusion_strategy": "attention",
+        "head_hidden_dim": 128,
+        "set_transformer_block": "sab"
+      },
+      "epoch_mean": 12,
+      "test_metrics": {
+        "size": {
+          "n_samples": 84,
+          "mse": 0.15145313313910996,
+          "rmse": 0.3891697998806048,
+          "mae": 0.2744193644750686,
+          "r2": -0.09948877551085711
+        },
+        "delivery": {
+          "n_samples": 58,
+          "mse": 0.9378577638427679,
+          "rmse": 0.9684305673835207,
+          "mae": 0.6135335185661398,
+          "r2": 0.09310702461562304
+        },
+        "pdi": {
+          "n_samples": 84,
+          "accuracy": 0.6071428571428571,
+          "precision": 0.4863247863247863,
+          "recall": 0.7491039426523297,
+          "f1": 0.5019607843137255
+        },
+        "ee": {
+          "n_samples": 84,
+          "accuracy": 0.6785714285714286,
+          "precision": 0.6324786324786325,
+          "recall": 0.7021367521367522,
+          "f1": 0.6488452445899254
+        },
+        "toxic": {
+          "n_samples": 59,
+          "accuracy": 0.9661016949152542,
+          "precision": 0.8,
+          "recall": 0.9821428571428572,
+          "f1": 0.8659090909090909
+        },
+        "biodist": {
+          "n_samples": 58,
+          "kl_divergence": 0.39116454524118993,
+          "js_divergence": 0.1065009133663494
+        }
+      }
+    }
+  ],
+  "summary_stats": {
+    "size": {
+      "mse_mean": 0.2768235036159284,
+      "mse_std": 0.1346583260723207,
+      "rmse_mean": 0.508798491275714,
+      "rmse_std": 0.13396864891266763,
+      "mae_mean": 0.3427187344475763,
+      "mae_std": 0.06271329639776796,
+      "r2_mean": -0.9110994850991861,
+      "r2_std": 1.3609378589681
+    },
+    "delivery": {
+      "mse_mean": 0.834075853950265,
+      "mse_std": 0.13581578656540713,
+      "rmse_mean": 0.9102708371721865,
+      "rmse_std": 0.0740463162089242,
+      "mae_mean": 0.634870878942305,
+      "mae_std": 0.04810398332825185,
+      "r2_mean": 0.14309063048291837,
+      "r2_std": 0.06678544282427759
+    },
+    "pdi": {
+      "accuracy_mean": 0.573809523809524,
+      "accuracy_std": 0.07315376902732006,
+      "precision_mean": 0.40644301832628565,
+      "precision_std": 0.06081224559746048,
+      "recall_mean": 0.591894166836878,
+      "recall_std": 0.15212645465643837,
+      "f1_mean": 0.40503518503422653,
+      "f1_std": 0.07270554354315185
+    },
+    "ee": {
+      "accuracy_mean": 0.65,
+      "accuracy_std": 0.05460640448180816,
+      "precision_mean": 0.5988843624524457,
+      "precision_std": 0.055997857124168736,
+      "recall_mean": 0.6355952320322067,
+      "recall_std": 0.07902933872282923,
+      "f1_mean": 0.5997699900861396,
+      "f1_std": 0.06505269720858664
+    },
+    "toxic": {
+      "accuracy_mean": 0.9500396342534485,
+      "accuracy_std": 0.014290090778642328,
+      "precision_mean": 0.7457142857142857,
+      "precision_std": 0.03149343955006944,
+      "recall_mean": 0.9737706334802525,
+      "recall_std": 0.007575117962473199,
+      "f1_mean": 0.81483431023254,
+      "f1_std": 0.0312513648447909
+    },
+    "biodist": {
+      "kl_divergence_mean": 0.42024315728866957,
+      "kl_divergence_std": 0.04170766696370389,
+      "js_divergence_mean": 0.10843433388487847,
+      "js_divergence_std": 0.010645869413863087
+    }
+  }
+}

models/abl/both/20260617_232841/outer_fold_0/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.12930599750984867,
+  "lr": 0.0003183893957684731,
+  "weight_decay": 0.0011797820449629415,
+  "backbone_lr_ratio": 0.35598913444921026,
+  "moe_n_experts": 2,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 2,
+  "llm_lora_r": 8,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/outer_fold_0/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 19, 21, 22, 23, 24, 25, 26, 27, 28, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 52, 53, 54, 55, 56, 57, 58, 61, 62, 63, 64, 65, 66, 67, 68, 71, 72, 73, 75, 76, 77, 80, 81, 85, 86, 88, 89, 90, 92, 94, 95, 96, 98, 99, 100, 101, 103, 104, 105, 106, 107, 108, 109, 111, 113, 114, 115, 116, 117, 118, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 133, 134, 135, 136, 138, 139, 141, 144, 145, 146, 147, 148, 149, 150, 151, 153, 154, 155, 156, 157, 158, 159, 160, 163, 164, 167, 168, 169, 170, 171, 172, 175, 176, 178, 179, 180, 182, 183, 184, 185, 188, 189, 190, 191, 192, 193, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 206, 208, 210, 211, 212, 213, 216, 217, 218, 219, 220, 222, 224, 225, 227, 228, 229, 230, 231, 232, 233, 234, 236, 237, 239, 241, 242, 244, 245, 247, 248, 249, 250, 252, 253, 255, 256, 257, 258, 259, 260, 261, 262, 265, 267, 268, 269, 270, 271, 272, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 285, 286, 287, 288, 291, 292, 294, 296, 297, 298, 299, 300, 301, 303, 304, 306, 307, 308, 309, 310, 311, 312, 313, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329, 330, 332, 333, 335, 336, 337, 338, 339, 340, 342, 343, 344, 345, 346, 347, 348, 349, 351, 352, 354, 355, 356, 357, 358, 359, 360, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 374, 375, 377, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389, 390, 391, 393, 395, 396, 397, 398, 399, 400, 401, 403, 404, 405, 406, 407, 408, 409, 411, 412, 413, 414, 415, 417, 418, 419], "outer_test_idx": [7, 16, 20, 29, 51, 59, 60, 69, 70, 74, 78, 79, 82, 83, 84, 87, 91, 93, 97, 102, 110, 112, 119, 120, 121, 132, 137, 140, 142, 143, 152, 161, 162, 165, 166, 173, 174, 177, 181, 186, 187, 194, 205, 207, 209, 214, 215, 221, 223, 226, 235, 238, 240, 243, 246, 251, 254, 263, 264, 266, 273, 284, 289, 290, 293, 295, 302, 305, 314, 315, 331, 334, 341, 350, 353, 361, 373, 376, 378, 392, 394, 402, 410, 416]}

models/abl/both/20260617_232841/outer_fold_0/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.27316252507783767,
+    "rmse": 0.5226495241343262,
+    "mae": 0.268634506018765,
+    "r2": 0.09300240292875017
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.49318943105662066,
+    "rmse": 0.7022744698881063,
+    "mae": 0.5225495728182381,
+    "r2": 0.37312647214479966
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.75,
+    "precision": 0.4107142857142857,
+    "recall": 0.5977751756440282,
+    "f1": 0.447008547008547
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.7023809523809523,
+    "precision": 0.6658119658119658,
+    "recall": 0.6976190476190475,
+    "f1": 0.6690688363683327
+  },
+  "toxic": {
+    "n_samples": 58,
+    "accuracy": 0.9655172413793104,
+    "precision": 0.75,
+    "recall": 0.9821428571428572,
+    "f1": 0.8242424242424242
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.20575334902695963,
+    "js_divergence": 0.045069148869594025
+  }
+}

models/abl/both/20260617_232841/outer_fold_1/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.14959924395216218,
+  "lr": 0.00012401964324838103,
+  "weight_decay": 9.656428689903883e-05,
+  "backbone_lr_ratio": 0.32683834688760083,
+  "moe_n_experts": 4,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 1,
+  "llm_lora_r": 32,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/outer_fold_1/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 3, 5, 6, 7, 10, 11, 12, 13, 15, 16, 18, 19, 20, 21, 24, 26, 27, 29, 30, 31, 32, 33, 35, 36, 38, 39, 40, 42, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 65, 66, 67, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 97, 98, 99, 100, 101, 102, 103, 105, 106, 107, 108, 109, 110, 112, 114, 115, 118, 119, 120, 121, 123, 124, 125, 126, 128, 129, 130, 131, 132, 136, 137, 140, 141, 142, 143, 144, 145, 146, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 161, 162, 164, 165, 166, 167, 168, 169, 171, 172, 173, 174, 175, 176, 177, 178, 179, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 194, 195, 196, 197, 198, 199, 201, 202, 204, 205, 206, 207, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 225, 226, 228, 229, 231, 232, 233, 234, 235, 236, 237, 238, 240, 242, 243, 244, 246, 247, 248, 250, 251, 252, 254, 255, 256, 258, 259, 260, 261, 262, 263, 264, 266, 268, 269, 270, 271, 273, 274, 275, 276, 277, 278, 279, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 295, 296, 297, 300, 301, 302, 303, 304, 305, 306, 307, 308, 310, 311, 313, 314, 315, 316, 317, 318, 319, 321, 322, 323, 324, 325, 327, 328, 329, 330, 331, 333, 334, 335, 337, 338, 339, 340, 341, 342, 343, 344, 345, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 361, 362, 364, 365, 366, 367, 368, 369, 373, 374, 375, 376, 378, 379, 380, 381, 382, 384, 387, 388, 389, 391, 392, 393, 394, 396, 397, 398, 399, 402, 403, 404, 405, 406, 408, 409, 410, 411, 412, 413, 414, 416, 417, 419], "outer_test_idx": [1, 2, 4, 8, 9, 14, 17, 22, 23, 25, 28, 34, 37, 41, 43, 45, 63, 64, 68, 85, 96, 104, 111, 113, 116, 117, 122, 127, 133, 134, 135, 138, 139, 147, 160, 163, 170, 180, 191, 192, 193, 200, 203, 208, 224, 227, 230, 239, 241, 245, 249, 253, 257, 265, 267, 272, 280, 294, 298, 299, 309, 312, 320, 326, 332, 336, 346, 347, 360, 363, 370, 371, 372, 377, 383, 385, 386, 390, 395, 400, 401, 407, 415, 418]}

models/abl/both/20260617_232841/outer_fold_1/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.11199287760383045,
+    "rmse": 0.3346533693298641,
+    "mae": 0.24370252518426805,
+    "r2": -0.8553464014912291
+  },
+  "delivery": {
+    "n_samples": 61,
+    "mse": 0.9945337433308287,
+    "rmse": 0.9972631264269368,
+    "mae": 0.637184505488296,
+    "r2": 0.24720965117371863
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5833333333333334,
+    "precision": 0.2606719367588933,
+    "recall": 0.24206349206349206,
+    "f1": 0.25038520801232667
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.5357142857142857,
+    "precision": 0.4737424924924925,
+    "recall": 0.47331240188383045,
+    "f1": 0.46359870390878144
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9508196721311475,
+    "precision": 0.75,
+    "recall": 0.9741379310344828,
+    "f1": 0.8200589970501475
+  },
+  "biodist": {
+    "n_samples": 61,
+    "kl_divergence": 0.42904510824573017,
+    "js_divergence": 0.11140172525390164
+  }
+}

models/abl/both/20260617_232841/outer_fold_2/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.12302210383900847,
+  "lr": 0.00015510064533248203,
+  "weight_decay": 0.0037340196924722196,
+  "backbone_lr_ratio": 0.3036603338505521,
+  "moe_n_experts": 4,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 1,
+  "llm_lora_r": 16,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/outer_fold_2/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 11, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 32, 33, 34, 36, 37, 38, 39, 40, 41, 43, 44, 45, 47, 48, 49, 50, 51, 55, 56, 57, 58, 59, 60, 61, 63, 64, 65, 68, 69, 70, 72, 74, 76, 78, 79, 80, 82, 83, 84, 85, 86, 87, 88, 91, 93, 94, 95, 96, 97, 100, 101, 102, 103, 104, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 142, 143, 145, 147, 148, 149, 150, 151, 152, 153, 156, 157, 158, 159, 160, 161, 162, 163, 165, 166, 167, 168, 170, 171, 172, 173, 174, 175, 177, 179, 180, 181, 183, 184, 185, 186, 187, 188, 191, 192, 193, 194, 196, 197, 199, 200, 202, 203, 204, 205, 207, 208, 209, 210, 211, 212, 213, 214, 215, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 232, 233, 235, 236, 237, 238, 239, 240, 241, 242, 243, 245, 246, 249, 250, 251, 252, 253, 254, 255, 256, 257, 259, 260, 261, 263, 264, 265, 266, 267, 268, 269, 271, 272, 273, 274, 275, 276, 280, 281, 282, 283, 284, 285, 287, 289, 290, 293, 294, 295, 296, 297, 298, 299, 300, 302, 303, 305, 307, 308, 309, 310, 311, 312, 313, 314, 315, 317, 318, 320, 321, 322, 323, 324, 326, 327, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 358, 360, 361, 363, 364, 365, 366, 368, 369, 370, 371, 372, 373, 376, 377, 378, 379, 380, 381, 382, 383, 385, 386, 387, 388, 389, 390, 392, 394, 395, 398, 399, 400, 401, 402, 404, 405, 407, 408, 409, 410, 412, 413, 414, 415, 416, 417, 418], "outer_test_idx": [10, 12, 15, 31, 35, 42, 46, 52, 53, 54, 62, 66, 67, 71, 73, 75, 77, 81, 89, 90, 92, 98, 99, 105, 123, 124, 141, 144, 146, 154, 155, 164, 169, 176, 178, 182, 189, 190, 195, 198, 201, 206, 216, 217, 218, 219, 231, 234, 244, 247, 248, 258, 262, 270, 277, 278, 279, 286, 288, 291, 292, 301, 304, 306, 316, 319, 325, 328, 343, 357, 359, 362, 367, 374, 375, 384, 391, 393, 396, 397, 403, 406, 411, 419]}

models/abl/both/20260617_232841/outer_fold_2/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.13105160735348353,
+    "rmse": 0.36201050724182515,
+    "mae": 0.29236553396497456,
+    "r2": -0.4041762321007929
+  },
+  "delivery": {
+    "n_samples": 60,
+    "mse": 0.6170164560338233,
+    "rmse": 0.7855039503616918,
+    "mae": 0.518595163594,
+    "r2": 0.178644153319804
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5714285714285714,
+    "precision": 0.3343023255813954,
+    "recall": 0.5261136712749617,
+    "f1": 0.34470899470899463
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6904761904761905,
+    "precision": 0.6311332192269291,
+    "recall": 0.6916433239962653,
+    "f1": 0.6470105449067581
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9344262295081968,
+    "precision": 0.7142857142857143,
+    "recall": 0.9655172413793103,
+    "f1": 0.7821428571428571
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.2793246003181208,
+    "js_divergence": 0.07360589998440034
+  }
+}

models/abl/both/20260617_232841/outer_fold_3/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.11061601330629428,
+  "lr": 0.0003491022701087641,
+  "weight_decay": 1.2696904821623371e-05,
+  "backbone_lr_ratio": 0.39277239063286945,
+  "moe_n_experts": 4,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 2,
+  "llm_lora_r": 32,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/outer_fold_3/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 4, 5, 7, 8, 9, 10, 11, 12, 14, 15, 16, 17, 20, 21, 22, 23, 24, 25, 27, 28, 29, 30, 31, 32, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 45, 46, 47, 51, 52, 53, 54, 55, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 73, 74, 75, 76, 77, 78, 79, 81, 82, 83, 84, 85, 86, 87, 89, 90, 91, 92, 93, 95, 96, 97, 98, 99, 100, 102, 104, 105, 106, 109, 110, 111, 112, 113, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 127, 129, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 146, 147, 148, 150, 151, 152, 153, 154, 155, 156, 158, 159, 160, 161, 162, 163, 164, 165, 166, 168, 169, 170, 172, 173, 174, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 189, 190, 191, 192, 193, 194, 195, 198, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 212, 213, 214, 215, 216, 217, 218, 219, 221, 223, 224, 226, 227, 228, 230, 231, 233, 234, 235, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 251, 252, 253, 254, 255, 256, 257, 258, 259, 262, 263, 264, 265, 266, 267, 268, 269, 270, 272, 273, 274, 275, 276, 277, 278, 279, 280, 282, 284, 286, 288, 289, 290, 291, 292, 293, 294, 295, 297, 298, 299, 301, 302, 304, 305, 306, 307, 308, 309, 310, 312, 314, 315, 316, 317, 319, 320, 321, 325, 326, 327, 328, 331, 332, 334, 336, 340, 341, 343, 344, 346, 347, 348, 350, 352, 353, 357, 359, 360, 361, 362, 363, 364, 366, 367, 368, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 382, 383, 384, 385, 386, 387, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 400, 401, 402, 403, 404, 406, 407, 410, 411, 412, 414, 415, 416, 418, 419], "outer_test_idx": [3, 6, 13, 18, 19, 26, 33, 44, 48, 49, 50, 56, 57, 72, 80, 88, 94, 101, 103, 107, 108, 114, 126, 128, 130, 145, 149, 157, 167, 171, 175, 188, 196, 197, 199, 211, 220, 222, 225, 229, 232, 236, 250, 260, 261, 271, 281, 283, 285, 287, 296, 300, 303, 311, 313, 318, 322, 323, 324, 329, 330, 333, 335, 337, 338, 339, 342, 345, 349, 351, 354, 355, 356, 358, 365, 369, 381, 388, 399, 405, 408, 409, 413, 417]}

models/abl/both/20260617_232841/outer_fold_3/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.2974639059494916,
+    "rmse": 0.5454025173662949,
+    "mae": 0.2553033943635872,
+    "r2": 0.08722103632340716
+  },
+  "delivery": {
+    "n_samples": 59,
+    "mse": 0.72832486531675,
+    "rmse": 0.8534195130864715,
+    "mae": 0.5703363718970095,
+    "r2": 0.29198169797506535
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.6785714285714286,
+    "precision": 0.3663025210084033,
+    "recall": 0.5624039938556068,
+    "f1": 0.3811802616390919
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6666666666666666,
+    "precision": 0.5903252773977808,
+    "recall": 0.6201608848667672,
+    "f1": 0.59879647539222
+  },
+  "toxic": {
+    "n_samples": 60,
+    "accuracy": 0.9333333333333333,
+    "precision": 0.7142857142857143,
+    "recall": 0.9649122807017544,
+    "f1": 0.7818181818181817
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.4182398191784831,
+    "js_divergence": 0.10764099439445415
+  }
+}

models/abl/both/20260617_232841/outer_fold_4/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.41353294032230026,
+  "lr": 0.00018606616740103076,
+  "weight_decay": 9.912252717813433e-05,
+  "backbone_lr_ratio": 0.32820993966799533,
+  "moe_n_experts": 4,
+  "moe_top_k": 1,
+  "moe_expert_hidden_mult": 2,
+  "llm_lora_r": 16,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/outer_fold_4/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [1, 2, 3, 4, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17, 18, 19, 20, 22, 23, 25, 26, 28, 29, 31, 33, 34, 35, 37, 41, 42, 43, 44, 45, 46, 48, 49, 50, 51, 52, 53, 54, 56, 57, 59, 60, 62, 63, 64, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 77, 78, 79, 80, 81, 82, 83, 84, 85, 87, 88, 89, 90, 91, 92, 93, 94, 96, 97, 98, 99, 101, 102, 103, 104, 105, 107, 108, 110, 111, 112, 113, 114, 116, 117, 119, 120, 121, 122, 123, 124, 126, 127, 128, 130, 132, 133, 134, 135, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 149, 152, 154, 155, 157, 160, 161, 162, 163, 164, 165, 166, 167, 169, 170, 171, 173, 174, 175, 176, 177, 178, 180, 181, 182, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 205, 206, 207, 208, 209, 211, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 234, 235, 236, 238, 239, 240, 241, 243, 244, 245, 246, 247, 248, 249, 250, 251, 253, 254, 257, 258, 260, 261, 262, 263, 264, 265, 266, 267, 270, 271, 272, 273, 277, 278, 279, 280, 281, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 298, 299, 300, 301, 302, 303, 304, 305, 306, 309, 311, 312, 313, 314, 315, 316, 318, 319, 320, 322, 323, 324, 325, 326, 328, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 341, 342, 343, 345, 346, 347, 349, 350, 351, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 365, 367, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 381, 383, 384, 385, 386, 388, 390, 391, 392, 393, 394, 395, 396, 397, 399, 400, 401, 402, 403, 405, 406, 407, 408, 409, 410, 411, 413, 415, 416, 417, 418, 419], "outer_test_idx": [0, 5, 11, 21, 24, 27, 30, 32, 36, 38, 39, 40, 47, 55, 58, 61, 65, 76, 86, 95, 100, 106, 109, 115, 118, 125, 129, 131, 136, 148, 150, 151, 153, 156, 158, 159, 168, 172, 179, 183, 184, 185, 202, 204, 210, 212, 213, 228, 233, 237, 242, 252, 255, 256, 259, 268, 269, 274, 275, 276, 282, 297, 307, 308, 310, 317, 321, 327, 340, 344, 348, 352, 364, 366, 368, 379, 380, 382, 387, 389, 398, 404, 412, 414]}

models/abl/both/20260617_232841/outer_fold_4/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.2385472153185827,
+    "rmse": 0.4884129557235175,
+    "mae": 0.37417884383882793,
+    "r2": -0.7317567503291509
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.7980907160636062,
+    "rmse": 0.8933592312522473,
+    "mae": 0.6033944039252298,
+    "r2": 0.2282594525293985
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5952380952380952,
+    "precision": 0.33809196980096085,
+    "recall": 0.5263056835637481,
+    "f1": 0.3583862620559868
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.7261904761904762,
+    "precision": 0.6831659493065518,
+    "recall": 0.7594017094017094,
+    "f1": 0.7065123352302344
+  },
+  "toxic": {
+    "n_samples": 59,
+    "accuracy": 0.9661016949152542,
+    "precision": 0.8,
+    "recall": 0.9821428571428572,
+    "f1": 0.8659090909090909
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.4050166777635646,
+    "js_divergence": 0.10143992174009567
+  }
+}

models/abl/both/20260617_232841/repeat_1/outer_fold_0/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.12930599750984867,
+  "lr": 0.0003183893957684731,
+  "weight_decay": 0.0011797820449629415,
+  "backbone_lr_ratio": 0.35598913444921026,
+  "moe_n_experts": 2,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 2,
+  "llm_lora_r": 8,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/repeat_1/outer_fold_0/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 19, 21, 22, 23, 24, 25, 26, 27, 28, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 52, 53, 54, 55, 56, 57, 58, 61, 62, 63, 64, 65, 66, 67, 68, 71, 72, 73, 75, 76, 77, 80, 81, 85, 86, 88, 89, 90, 92, 94, 95, 96, 98, 99, 100, 101, 103, 104, 105, 106, 107, 108, 109, 111, 113, 114, 115, 116, 117, 118, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 133, 134, 135, 136, 138, 139, 141, 144, 145, 146, 147, 148, 149, 150, 151, 153, 154, 155, 156, 157, 158, 159, 160, 163, 164, 167, 168, 169, 170, 171, 172, 175, 176, 178, 179, 180, 182, 183, 184, 185, 188, 189, 190, 191, 192, 193, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 206, 208, 210, 211, 212, 213, 216, 217, 218, 219, 220, 222, 224, 225, 227, 228, 229, 230, 231, 232, 233, 234, 236, 237, 239, 241, 242, 244, 245, 247, 248, 249, 250, 252, 253, 255, 256, 257, 258, 259, 260, 261, 262, 265, 267, 268, 269, 270, 271, 272, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 285, 286, 287, 288, 291, 292, 294, 296, 297, 298, 299, 300, 301, 303, 304, 306, 307, 308, 309, 310, 311, 312, 313, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329, 330, 332, 333, 335, 336, 337, 338, 339, 340, 342, 343, 344, 345, 346, 347, 348, 349, 351, 352, 354, 355, 356, 357, 358, 359, 360, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 374, 375, 377, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389, 390, 391, 393, 395, 396, 397, 398, 399, 400, 401, 403, 404, 405, 406, 407, 408, 409, 411, 412, 413, 414, 415, 417, 418, 419], "outer_test_idx": [7, 16, 20, 29, 51, 59, 60, 69, 70, 74, 78, 79, 82, 83, 84, 87, 91, 93, 97, 102, 110, 112, 119, 120, 121, 132, 137, 140, 142, 143, 152, 161, 162, 165, 166, 173, 174, 177, 181, 186, 187, 194, 205, 207, 209, 214, 215, 221, 223, 226, 235, 238, 240, 243, 246, 251, 254, 263, 264, 266, 273, 284, 289, 290, 293, 295, 302, 305, 314, 315, 331, 334, 341, 350, 353, 361, 373, 376, 378, 392, 394, 402, 410, 416]}

models/abl/both/20260617_232841/repeat_1/outer_fold_0/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.2816563970171338,
+    "rmse": 0.5307131023605257,
+    "mae": 0.27628492734518395,
+    "r2": 0.06479970039267835
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.5807003895647976,
+    "rmse": 0.7620370001284699,
+    "mae": 0.564853833378132,
+    "r2": 0.26189476312686466
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.7857142857142857,
+    "precision": 0.42781279178338005,
+    "recall": 0.5944574551131928,
+    "f1": 0.4533531409168081
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6547619047619048,
+    "precision": 0.6603426956368134,
+    "recall": 0.7080952380952382,
+    "f1": 0.6472175061309188
+  },
+  "toxic": {
+    "n_samples": 58,
+    "accuracy": 0.9827586206896551,
+    "precision": 0.9912280701754386,
+    "recall": 0.75,
+    "f1": 0.8289085545722714
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.2790682434585921,
+    "js_divergence": 0.06757546969442413
+  }
+}

models/abl/both/20260617_232841/repeat_1/outer_fold_1/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.14959924395216218,
+  "lr": 0.00012401964324838103,
+  "weight_decay": 9.656428689903883e-05,
+  "backbone_lr_ratio": 0.32683834688760083,
+  "moe_n_experts": 4,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 1,
+  "llm_lora_r": 32,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/repeat_1/outer_fold_1/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 3, 5, 6, 7, 10, 11, 12, 13, 15, 16, 18, 19, 20, 21, 24, 26, 27, 29, 30, 31, 32, 33, 35, 36, 38, 39, 40, 42, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 65, 66, 67, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 97, 98, 99, 100, 101, 102, 103, 105, 106, 107, 108, 109, 110, 112, 114, 115, 118, 119, 120, 121, 123, 124, 125, 126, 128, 129, 130, 131, 132, 136, 137, 140, 141, 142, 143, 144, 145, 146, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 161, 162, 164, 165, 166, 167, 168, 169, 171, 172, 173, 174, 175, 176, 177, 178, 179, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 194, 195, 196, 197, 198, 199, 201, 202, 204, 205, 206, 207, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 225, 226, 228, 229, 231, 232, 233, 234, 235, 236, 237, 238, 240, 242, 243, 244, 246, 247, 248, 250, 251, 252, 254, 255, 256, 258, 259, 260, 261, 262, 263, 264, 266, 268, 269, 270, 271, 273, 274, 275, 276, 277, 278, 279, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 295, 296, 297, 300, 301, 302, 303, 304, 305, 306, 307, 308, 310, 311, 313, 314, 315, 316, 317, 318, 319, 321, 322, 323, 324, 325, 327, 328, 329, 330, 331, 333, 334, 335, 337, 338, 339, 340, 341, 342, 343, 344, 345, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 361, 362, 364, 365, 366, 367, 368, 369, 373, 374, 375, 376, 378, 379, 380, 381, 382, 384, 387, 388, 389, 391, 392, 393, 394, 396, 397, 398, 399, 402, 403, 404, 405, 406, 408, 409, 410, 411, 412, 413, 414, 416, 417, 419], "outer_test_idx": [1, 2, 4, 8, 9, 14, 17, 22, 23, 25, 28, 34, 37, 41, 43, 45, 63, 64, 68, 85, 96, 104, 111, 113, 116, 117, 122, 127, 133, 134, 135, 138, 139, 147, 160, 163, 170, 180, 191, 192, 193, 200, 203, 208, 224, 227, 230, 239, 241, 245, 249, 253, 257, 265, 267, 272, 280, 294, 298, 299, 309, 312, 320, 326, 332, 336, 346, 347, 360, 363, 370, 371, 372, 377, 383, 385, 386, 390, 395, 400, 401, 407, 415, 418]}

models/abl/both/20260617_232841/repeat_1/outer_fold_1/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.08666696329506099,
+    "rmse": 0.2943925326754417,
+    "mae": 0.22798642090388707,
+    "r2": -0.4357809346276156
+  },
+  "delivery": {
+    "n_samples": 61,
+    "mse": 0.9690008398737997,
+    "rmse": 0.9843784027871597,
+    "mae": 0.626403434202075,
+    "r2": 0.26653621845095465
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.4880952380952381,
+    "precision": 0.27691029900332226,
+    "recall": 0.2579365079365079,
+    "f1": 0.24170918367346939
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.5357142857142857,
+    "precision": 0.4703263535182041,
+    "recall": 0.47331240188383045,
+    "f1": 0.46348793258357346
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9508196721311475,
+    "precision": 0.75,
+    "recall": 0.9741379310344828,
+    "f1": 0.8200589970501475
+  },
+  "biodist": {
+    "n_samples": 61,
+    "kl_divergence": 0.4359101867730315,
+    "js_divergence": 0.11254492304386336
+  }
+}

models/abl/both/20260617_232841/repeat_1/outer_fold_2/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.12302210383900847,
+  "lr": 0.00015510064533248203,
+  "weight_decay": 0.0037340196924722196,
+  "backbone_lr_ratio": 0.3036603338505521,
+  "moe_n_experts": 4,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 1,
+  "llm_lora_r": 16,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/repeat_1/outer_fold_2/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 11, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 32, 33, 34, 36, 37, 38, 39, 40, 41, 43, 44, 45, 47, 48, 49, 50, 51, 55, 56, 57, 58, 59, 60, 61, 63, 64, 65, 68, 69, 70, 72, 74, 76, 78, 79, 80, 82, 83, 84, 85, 86, 87, 88, 91, 93, 94, 95, 96, 97, 100, 101, 102, 103, 104, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 142, 143, 145, 147, 148, 149, 150, 151, 152, 153, 156, 157, 158, 159, 160, 161, 162, 163, 165, 166, 167, 168, 170, 171, 172, 173, 174, 175, 177, 179, 180, 181, 183, 184, 185, 186, 187, 188, 191, 192, 193, 194, 196, 197, 199, 200, 202, 203, 204, 205, 207, 208, 209, 210, 211, 212, 213, 214, 215, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 232, 233, 235, 236, 237, 238, 239, 240, 241, 242, 243, 245, 246, 249, 250, 251, 252, 253, 254, 255, 256, 257, 259, 260, 261, 263, 264, 265, 266, 267, 268, 269, 271, 272, 273, 274, 275, 276, 280, 281, 282, 283, 284, 285, 287, 289, 290, 293, 294, 295, 296, 297, 298, 299, 300, 302, 303, 305, 307, 308, 309, 310, 311, 312, 313, 314, 315, 317, 318, 320, 321, 322, 323, 324, 326, 327, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 358, 360, 361, 363, 364, 365, 366, 368, 369, 370, 371, 372, 373, 376, 377, 378, 379, 380, 381, 382, 383, 385, 386, 387, 388, 389, 390, 392, 394, 395, 398, 399, 400, 401, 402, 404, 405, 407, 408, 409, 410, 412, 413, 414, 415, 416, 417, 418], "outer_test_idx": [10, 12, 15, 31, 35, 42, 46, 52, 53, 54, 62, 66, 67, 71, 73, 75, 77, 81, 89, 90, 92, 98, 99, 105, 123, 124, 141, 144, 146, 154, 155, 164, 169, 176, 178, 182, 189, 190, 195, 198, 201, 206, 216, 217, 218, 219, 231, 234, 244, 247, 248, 258, 262, 270, 277, 278, 279, 286, 288, 291, 292, 301, 304, 306, 316, 319, 325, 328, 343, 357, 359, 362, 367, 374, 375, 384, 391, 393, 396, 397, 403, 406, 411, 419]}

models/abl/both/20260617_232841/repeat_1/outer_fold_2/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.13747360050361937,
+    "rmse": 0.3707743255723343,
+    "mae": 0.2868303656578064,
+    "r2": -0.4729858432626901
+  },
+  "delivery": {
+    "n_samples": 60,
+    "mse": 0.7104536138129312,
+    "rmse": 0.8428841046151785,
+    "mae": 0.5430462850956246,
+    "r2": 0.054263101423121185
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.6428571428571429,
+    "precision": 0.45827160493827157,
+    "recall": 0.7127496159754224,
+    "f1": 0.4944194569381975
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6666666666666666,
+    "precision": 0.6060343270099368,
+    "recall": 0.6573295985060691,
+    "f1": 0.6198067632850242
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9344262295081968,
+    "precision": 0.7142857142857143,
+    "recall": 0.9655172413793103,
+    "f1": 0.7821428571428571
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.28003378134681434,
+    "js_divergence": 0.0731604987811758
+  }
+}

models/abl/both/20260617_232841/repeat_1/outer_fold_3/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.11061601330629428,
+  "lr": 0.0003491022701087641,
+  "weight_decay": 1.2696904821623371e-05,
+  "backbone_lr_ratio": 0.39277239063286945,
+  "moe_n_experts": 4,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 2,
+  "llm_lora_r": 32,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/repeat_1/outer_fold_3/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 4, 5, 7, 8, 9, 10, 11, 12, 14, 15, 16, 17, 20, 21, 22, 23, 24, 25, 27, 28, 29, 30, 31, 32, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 45, 46, 47, 51, 52, 53, 54, 55, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 73, 74, 75, 76, 77, 78, 79, 81, 82, 83, 84, 85, 86, 87, 89, 90, 91, 92, 93, 95, 96, 97, 98, 99, 100, 102, 104, 105, 106, 109, 110, 111, 112, 113, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 127, 129, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 146, 147, 148, 150, 151, 152, 153, 154, 155, 156, 158, 159, 160, 161, 162, 163, 164, 165, 166, 168, 169, 170, 172, 173, 174, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 189, 190, 191, 192, 193, 194, 195, 198, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 212, 213, 214, 215, 216, 217, 218, 219, 221, 223, 224, 226, 227, 228, 230, 231, 233, 234, 235, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 251, 252, 253, 254, 255, 256, 257, 258, 259, 262, 263, 264, 265, 266, 267, 268, 269, 270, 272, 273, 274, 275, 276, 277, 278, 279, 280, 282, 284, 286, 288, 289, 290, 291, 292, 293, 294, 295, 297, 298, 299, 301, 302, 304, 305, 306, 307, 308, 309, 310, 312, 314, 315, 316, 317, 319, 320, 321, 325, 326, 327, 328, 331, 332, 334, 336, 340, 341, 343, 344, 346, 347, 348, 350, 352, 353, 357, 359, 360, 361, 362, 363, 364, 366, 367, 368, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 382, 383, 384, 385, 386, 387, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 400, 401, 402, 403, 404, 406, 407, 410, 411, 412, 414, 415, 416, 418, 419], "outer_test_idx": [3, 6, 13, 18, 19, 26, 33, 44, 48, 49, 50, 56, 57, 72, 80, 88, 94, 101, 103, 107, 108, 114, 126, 128, 130, 145, 149, 157, 167, 171, 175, 188, 196, 197, 199, 211, 220, 222, 225, 229, 232, 236, 250, 260, 261, 271, 281, 283, 285, 287, 296, 300, 303, 311, 313, 318, 322, 323, 324, 329, 330, 333, 335, 337, 338, 339, 342, 345, 349, 351, 354, 355, 356, 358, 365, 369, 381, 388, 399, 405, 408, 409, 413, 417]}

models/abl/both/20260617_232841/repeat_1/outer_fold_3/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.32669708395989955,
+    "rmse": 0.5715742156184965,
+    "mae": 0.31596550309514426,
+    "r2": -0.0024820483050997932
+  },
+  "delivery": {
+    "n_samples": 59,
+    "mse": 0.7029380513850005,
+    "rmse": 0.8384140095352657,
+    "mae": 0.5650059677881457,
+    "r2": 0.31666069734708746
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.6785714285714286,
+    "precision": 0.3406323185011709,
+    "recall": 0.5230414746543779,
+    "f1": 0.3490853658536585
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6547619047619048,
+    "precision": 0.5888888888888889,
+    "recall": 0.6327300150829562,
+    "f1": 0.5970695970695971
+  },
+  "toxic": {
+    "n_samples": 60,
+    "accuracy": 0.9333333333333333,
+    "precision": 0.7142857142857143,
+    "recall": 0.9649122807017544,
+    "f1": 0.7818181818181817
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.3455956570999495,
+    "js_divergence": 0.08991455654751294
+  }
+}

models/abl/both/20260617_232841/repeat_1/outer_fold_4/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.41353294032230026,
+  "lr": 0.00018606616740103076,
+  "weight_decay": 9.912252717813433e-05,
+  "backbone_lr_ratio": 0.32820993966799533,
+  "moe_n_experts": 4,
+  "moe_top_k": 1,
+  "moe_expert_hidden_mult": 2,
+  "llm_lora_r": 16,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/repeat_1/outer_fold_4/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [1, 2, 3, 4, 6, 7, 8, 9, 10, 12, 13, 14, 15, 16, 17, 18, 19, 20, 22, 23, 25, 26, 28, 29, 31, 33, 34, 35, 37, 41, 42, 43, 44, 45, 46, 48, 49, 50, 51, 52, 53, 54, 56, 57, 59, 60, 62, 63, 64, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 77, 78, 79, 80, 81, 82, 83, 84, 85, 87, 88, 89, 90, 91, 92, 93, 94, 96, 97, 98, 99, 101, 102, 103, 104, 105, 107, 108, 110, 111, 112, 113, 114, 116, 117, 119, 120, 121, 122, 123, 124, 126, 127, 128, 130, 132, 133, 134, 135, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 149, 152, 154, 155, 157, 160, 161, 162, 163, 164, 165, 166, 167, 169, 170, 171, 173, 174, 175, 176, 177, 178, 180, 181, 182, 186, 187, 188, 189, 190, 191, 192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 205, 206, 207, 208, 209, 211, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 234, 235, 236, 238, 239, 240, 241, 243, 244, 245, 246, 247, 248, 249, 250, 251, 253, 254, 257, 258, 260, 261, 262, 263, 264, 265, 266, 267, 270, 271, 272, 273, 277, 278, 279, 280, 281, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 294, 295, 296, 298, 299, 300, 301, 302, 303, 304, 305, 306, 309, 311, 312, 313, 314, 315, 316, 318, 319, 320, 322, 323, 324, 325, 326, 328, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 341, 342, 343, 345, 346, 347, 349, 350, 351, 353, 354, 355, 356, 357, 358, 359, 360, 361, 362, 363, 365, 367, 369, 370, 371, 372, 373, 374, 375, 376, 377, 378, 381, 383, 384, 385, 386, 388, 390, 391, 392, 393, 394, 395, 396, 397, 399, 400, 401, 402, 403, 405, 406, 407, 408, 409, 410, 411, 413, 415, 416, 417, 418, 419], "outer_test_idx": [0, 5, 11, 21, 24, 27, 30, 32, 36, 38, 39, 40, 47, 55, 58, 61, 65, 76, 86, 95, 100, 106, 109, 115, 118, 125, 129, 131, 136, 148, 150, 151, 153, 156, 158, 159, 168, 172, 179, 183, 184, 185, 202, 204, 210, 212, 213, 228, 233, 237, 242, 252, 255, 256, 259, 268, 269, 274, 275, 276, 282, 297, 307, 308, 310, 317, 321, 327, 340, 344, 348, 352, 364, 366, 368, 379, 380, 382, 387, 389, 398, 404, 412, 414]}

models/abl/both/20260617_232841/repeat_1/outer_fold_4/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.2498053256210641,
+    "rmse": 0.49980528770818755,
+    "mae": 0.4012378028460911,
+    "r2": -0.8134860989037498
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.8713336067211406,
+    "rmse": 0.9334525198000917,
+    "mae": 0.6369200194704121,
+    "r2": 0.1574347863646688
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.6190476190476191,
+    "precision": 0.32608695652173914,
+    "recall": 0.510752688172043,
+    "f1": 0.350912975912976
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6785714285714286,
+    "precision": 0.6324786324786325,
+    "recall": 0.7021367521367522,
+    "f1": 0.6488452445899254
+  },
+  "toxic": {
+    "n_samples": 59,
+    "accuracy": 0.9661016949152542,
+    "precision": 0.8,
+    "recall": 0.9821428571428572,
+    "f1": 0.8659090909090909
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.3635802403548602,
+    "js_divergence": 0.09224357999051915
+  }
+}

models/abl/both/20260617_232841/repeat_2/outer_fold_0/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.12930599750984867,
+  "lr": 0.0003183893957684731,
+  "weight_decay": 0.0011797820449629415,
+  "backbone_lr_ratio": 0.35598913444921026,
+  "moe_n_experts": 2,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 2,
+  "llm_lora_r": 8,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/repeat_2/outer_fold_0/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 13, 14, 15, 17, 18, 19, 21, 22, 23, 24, 25, 26, 27, 28, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 52, 53, 54, 55, 56, 57, 58, 61, 62, 63, 64, 65, 66, 67, 68, 71, 72, 73, 75, 76, 77, 80, 81, 85, 86, 88, 89, 90, 92, 94, 95, 96, 98, 99, 100, 101, 103, 104, 105, 106, 107, 108, 109, 111, 113, 114, 115, 116, 117, 118, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 133, 134, 135, 136, 138, 139, 141, 144, 145, 146, 147, 148, 149, 150, 151, 153, 154, 155, 156, 157, 158, 159, 160, 163, 164, 167, 168, 169, 170, 171, 172, 175, 176, 178, 179, 180, 182, 183, 184, 185, 188, 189, 190, 191, 192, 193, 195, 196, 197, 198, 199, 200, 201, 202, 203, 204, 206, 208, 210, 211, 212, 213, 216, 217, 218, 219, 220, 222, 224, 225, 227, 228, 229, 230, 231, 232, 233, 234, 236, 237, 239, 241, 242, 244, 245, 247, 248, 249, 250, 252, 253, 255, 256, 257, 258, 259, 260, 261, 262, 265, 267, 268, 269, 270, 271, 272, 274, 275, 276, 277, 278, 279, 280, 281, 282, 283, 285, 286, 287, 288, 291, 292, 294, 296, 297, 298, 299, 300, 301, 303, 304, 306, 307, 308, 309, 310, 311, 312, 313, 316, 317, 318, 319, 320, 321, 322, 323, 324, 325, 326, 327, 328, 329, 330, 332, 333, 335, 336, 337, 338, 339, 340, 342, 343, 344, 345, 346, 347, 348, 349, 351, 352, 354, 355, 356, 357, 358, 359, 360, 362, 363, 364, 365, 366, 367, 368, 369, 370, 371, 372, 374, 375, 377, 379, 380, 381, 382, 383, 384, 385, 386, 387, 388, 389, 390, 391, 393, 395, 396, 397, 398, 399, 400, 401, 403, 404, 405, 406, 407, 408, 409, 411, 412, 413, 414, 415, 417, 418, 419], "outer_test_idx": [7, 16, 20, 29, 51, 59, 60, 69, 70, 74, 78, 79, 82, 83, 84, 87, 91, 93, 97, 102, 110, 112, 119, 120, 121, 132, 137, 140, 142, 143, 152, 161, 162, 165, 166, 173, 174, 177, 181, 186, 187, 194, 205, 207, 209, 214, 215, 221, 223, 226, 235, 238, 240, 243, 246, 251, 254, 263, 264, 266, 273, 284, 289, 290, 293, 295, 302, 305, 314, 315, 331, 334, 341, 350, 353, 361, 373, 376, 378, 392, 394, 402, 410, 416]}

models/abl/both/20260617_232841/repeat_2/outer_fold_0/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.3048058934964392,
+    "rmse": 0.5520922871191366,
+    "mae": 0.30447895555611115,
+    "r2": -0.012064934220566759
+  },
+  "delivery": {
+    "n_samples": 58,
+    "mse": 0.5277170018903071,
+    "rmse": 0.7264413272180398,
+    "mae": 0.553587457165122,
+    "r2": 0.32923984608630563
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.7261904761904762,
+    "precision": 0.3958333333333333,
+    "recall": 0.5817720530835285,
+    "f1": 0.43162393162393164
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.75,
+    "precision": 0.7124588290649997,
+    "recall": 0.7785714285714285,
+    "f1": 0.7248325686863893
+  },
+  "toxic": {
+    "n_samples": 58,
+    "accuracy": 0.9655172413793104,
+    "precision": 0.75,
+    "recall": 0.9821428571428572,
+    "f1": 0.8242424242424242
+  },
+  "biodist": {
+    "n_samples": 58,
+    "kl_divergence": 0.2720172018326682,
+    "js_divergence": 0.060785283409926956
+  }
+}

models/abl/both/20260617_232841/repeat_2/outer_fold_1/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.14959924395216218,
+  "lr": 0.00012401964324838103,
+  "weight_decay": 9.656428689903883e-05,
+  "backbone_lr_ratio": 0.32683834688760083,
+  "moe_n_experts": 4,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 1,
+  "llm_lora_r": 32,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/repeat_2/outer_fold_1/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 3, 5, 6, 7, 10, 11, 12, 13, 15, 16, 18, 19, 20, 21, 24, 26, 27, 29, 30, 31, 32, 33, 35, 36, 38, 39, 40, 42, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 65, 66, 67, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 97, 98, 99, 100, 101, 102, 103, 105, 106, 107, 108, 109, 110, 112, 114, 115, 118, 119, 120, 121, 123, 124, 125, 126, 128, 129, 130, 131, 132, 136, 137, 140, 141, 142, 143, 144, 145, 146, 148, 149, 150, 151, 152, 153, 154, 155, 156, 157, 158, 159, 161, 162, 164, 165, 166, 167, 168, 169, 171, 172, 173, 174, 175, 176, 177, 178, 179, 181, 182, 183, 184, 185, 186, 187, 188, 189, 190, 194, 195, 196, 197, 198, 199, 201, 202, 204, 205, 206, 207, 209, 210, 211, 212, 213, 214, 215, 216, 217, 218, 219, 220, 221, 222, 223, 225, 226, 228, 229, 231, 232, 233, 234, 235, 236, 237, 238, 240, 242, 243, 244, 246, 247, 248, 250, 251, 252, 254, 255, 256, 258, 259, 260, 261, 262, 263, 264, 266, 268, 269, 270, 271, 273, 274, 275, 276, 277, 278, 279, 281, 282, 283, 284, 285, 286, 287, 288, 289, 290, 291, 292, 293, 295, 296, 297, 300, 301, 302, 303, 304, 305, 306, 307, 308, 310, 311, 313, 314, 315, 316, 317, 318, 319, 321, 322, 323, 324, 325, 327, 328, 329, 330, 331, 333, 334, 335, 337, 338, 339, 340, 341, 342, 343, 344, 345, 348, 349, 350, 351, 352, 353, 354, 355, 356, 357, 358, 359, 361, 362, 364, 365, 366, 367, 368, 369, 373, 374, 375, 376, 378, 379, 380, 381, 382, 384, 387, 388, 389, 391, 392, 393, 394, 396, 397, 398, 399, 402, 403, 404, 405, 406, 408, 409, 410, 411, 412, 413, 414, 416, 417, 419], "outer_test_idx": [1, 2, 4, 8, 9, 14, 17, 22, 23, 25, 28, 34, 37, 41, 43, 45, 63, 64, 68, 85, 96, 104, 111, 113, 116, 117, 122, 127, 133, 134, 135, 138, 139, 147, 160, 163, 170, 180, 191, 192, 193, 200, 203, 208, 224, 227, 230, 239, 241, 245, 249, 253, 257, 265, 267, 272, 280, 294, 298, 299, 309, 312, 320, 326, 332, 336, 346, 347, 360, 363, 370, 371, 372, 377, 383, 385, 386, 390, 395, 400, 401, 407, 415, 418]}

models/abl/both/20260617_232841/repeat_2/outer_fold_1/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.10482404388180623,
+    "rmse": 0.32376541489449767,
+    "mae": 0.23482046808515275,
+    "r2": -0.7365828681878246
+  },
+  "delivery": {
+    "n_samples": 61,
+    "mse": 0.9723624200125729,
+    "rmse": 0.9860843878758921,
+    "mae": 0.6302215732084434,
+    "r2": 0.2639917446186245
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5952380952380952,
+    "precision": 0.28258620689655173,
+    "recall": 0.2698412698412698,
+    "f1": 0.2714159292035398
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.5476190476190477,
+    "precision": 0.48611111111111116,
+    "recall": 0.4884639170353456,
+    "f1": 0.4753962887837127
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9508196721311475,
+    "precision": 0.75,
+    "recall": 0.9741379310344828,
+    "f1": 0.8200589970501475
+  },
+  "biodist": {
+    "n_samples": 61,
+    "kl_divergence": 0.4600248432068658,
+    "js_divergence": 0.11821523045175072
+  }
+}

models/abl/both/20260617_232841/repeat_2/outer_fold_2/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.12302210383900847,
+  "lr": 0.00015510064533248203,
+  "weight_decay": 0.0037340196924722196,
+  "backbone_lr_ratio": 0.3036603338505521,
+  "moe_n_experts": 4,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 1,
+  "llm_lora_r": 16,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/repeat_2/outer_fold_2/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 11, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 32, 33, 34, 36, 37, 38, 39, 40, 41, 43, 44, 45, 47, 48, 49, 50, 51, 55, 56, 57, 58, 59, 60, 61, 63, 64, 65, 68, 69, 70, 72, 74, 76, 78, 79, 80, 82, 83, 84, 85, 86, 87, 88, 91, 93, 94, 95, 96, 97, 100, 101, 102, 103, 104, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 142, 143, 145, 147, 148, 149, 150, 151, 152, 153, 156, 157, 158, 159, 160, 161, 162, 163, 165, 166, 167, 168, 170, 171, 172, 173, 174, 175, 177, 179, 180, 181, 183, 184, 185, 186, 187, 188, 191, 192, 193, 194, 196, 197, 199, 200, 202, 203, 204, 205, 207, 208, 209, 210, 211, 212, 213, 214, 215, 220, 221, 222, 223, 224, 225, 226, 227, 228, 229, 230, 232, 233, 235, 236, 237, 238, 239, 240, 241, 242, 243, 245, 246, 249, 250, 251, 252, 253, 254, 255, 256, 257, 259, 260, 261, 263, 264, 265, 266, 267, 268, 269, 271, 272, 273, 274, 275, 276, 280, 281, 282, 283, 284, 285, 287, 289, 290, 293, 294, 295, 296, 297, 298, 299, 300, 302, 303, 305, 307, 308, 309, 310, 311, 312, 313, 314, 315, 317, 318, 320, 321, 322, 323, 324, 326, 327, 329, 330, 331, 332, 333, 334, 335, 336, 337, 338, 339, 340, 341, 342, 344, 345, 346, 347, 348, 349, 350, 351, 352, 353, 354, 355, 356, 358, 360, 361, 363, 364, 365, 366, 368, 369, 370, 371, 372, 373, 376, 377, 378, 379, 380, 381, 382, 383, 385, 386, 387, 388, 389, 390, 392, 394, 395, 398, 399, 400, 401, 402, 404, 405, 407, 408, 409, 410, 412, 413, 414, 415, 416, 417, 418], "outer_test_idx": [10, 12, 15, 31, 35, 42, 46, 52, 53, 54, 62, 66, 67, 71, 73, 75, 77, 81, 89, 90, 92, 98, 99, 105, 123, 124, 141, 144, 146, 154, 155, 164, 169, 176, 178, 182, 189, 190, 195, 198, 201, 206, 216, 217, 218, 219, 231, 234, 244, 247, 248, 258, 262, 270, 277, 278, 279, 286, 288, 291, 292, 301, 304, 306, 316, 319, 325, 328, 343, 357, 359, 362, 367, 374, 375, 384, 391, 393, 396, 397, 403, 406, 411, 419]}

models/abl/both/20260617_232841/repeat_2/outer_fold_2/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 84,
+    "mse": 0.12549337649644715,
+    "rmse": 0.35425044318454585,
+    "mae": 0.2802231794311887,
+    "r2": -0.34462155879619094
+  },
+  "delivery": {
+    "n_samples": 60,
+    "mse": 0.6592902402854742,
+    "rmse": 0.8119668960527111,
+    "mae": 0.5348013029200956,
+    "r2": 0.12237041942365745
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.5119047619047619,
+    "precision": 0.4135659661975451,
+    "recall": 0.6746031746031745,
+    "f1": 0.4122222222222222
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6547619047619048,
+    "precision": 0.5972649572649572,
+    "recall": 0.6507936507936507,
+    "f1": 0.6069243156199677
+  },
+  "toxic": {
+    "n_samples": 61,
+    "accuracy": 0.9344262295081968,
+    "precision": 0.7142857142857143,
+    "recall": 0.9655172413793103,
+    "f1": 0.7821428571428571
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.2563386788633516,
+    "js_divergence": 0.06872839042316117
+  }
+}

models/abl/both/20260617_232841/repeat_2/outer_fold_3/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.11061601330629428,
+  "lr": 0.0003491022701087641,
+  "weight_decay": 1.2696904821623371e-05,
+  "backbone_lr_ratio": 0.39277239063286945,
+  "moe_n_experts": 4,
+  "moe_top_k": 2,
+  "moe_expert_hidden_mult": 2,
+  "llm_lora_r": 32,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}

models/abl/both/20260617_232841/repeat_2/outer_fold_3/splits.json

@@ -0,0 +1 @@
+{"outer_train_idx": [0, 1, 2, 4, 5, 7, 8, 9, 10, 11, 12, 14, 15, 16, 17, 20, 21, 22, 23, 24, 25, 27, 28, 29, 30, 31, 32, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 45, 46, 47, 51, 52, 53, 54, 55, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 73, 74, 75, 76, 77, 78, 79, 81, 82, 83, 84, 85, 86, 87, 89, 90, 91, 92, 93, 95, 96, 97, 98, 99, 100, 102, 104, 105, 106, 109, 110, 111, 112, 113, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 127, 129, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 146, 147, 148, 150, 151, 152, 153, 154, 155, 156, 158, 159, 160, 161, 162, 163, 164, 165, 166, 168, 169, 170, 172, 173, 174, 176, 177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 189, 190, 191, 192, 193, 194, 195, 198, 200, 201, 202, 203, 204, 205, 206, 207, 208, 209, 210, 212, 213, 214, 215, 216, 217, 218, 219, 221, 223, 224, 226, 227, 228, 230, 231, 233, 234, 235, 237, 238, 239, 240, 241, 242, 243, 244, 245, 246, 247, 248, 249, 251, 252, 253, 254, 255, 256, 257, 258, 259, 262, 263, 264, 265, 266, 267, 268, 269, 270, 272, 273, 274, 275, 276, 277, 278, 279, 280, 282, 284, 286, 288, 289, 290, 291, 292, 293, 294, 295, 297, 298, 299, 301, 302, 304, 305, 306, 307, 308, 309, 310, 312, 314, 315, 316, 317, 319, 320, 321, 325, 326, 327, 328, 331, 332, 334, 336, 340, 341, 343, 344, 346, 347, 348, 350, 352, 353, 357, 359, 360, 361, 362, 363, 364, 366, 367, 368, 370, 371, 372, 373, 374, 375, 376, 377, 378, 379, 380, 382, 383, 384, 385, 386, 387, 389, 390, 391, 392, 393, 394, 395, 396, 397, 398, 400, 401, 402, 403, 404, 406, 407, 410, 411, 412, 414, 415, 416, 418, 419], "outer_test_idx": [3, 6, 13, 18, 19, 26, 33, 44, 48, 49, 50, 56, 57, 72, 80, 88, 94, 101, 103, 107, 108, 114, 126, 128, 130, 145, 149, 157, 167, 171, 175, 188, 196, 197, 199, 211, 220, 222, 225, 229, 232, 236, 250, 260, 261, 271, 281, 283, 285, 287, 296, 300, 303, 311, 313, 318, 322, 323, 324, 329, 330, 333, 335, 337, 338, 339, 342, 345, 349, 351, 354, 355, 356, 358, 365, 369, 381, 388, 399, 405, 408, 409, 413, 417]}

models/abl/both/20260617_232841/repeat_2/outer_fold_3/test_metrics.json

@@ -0,0 +1,42 @@
+{
+  "size": {
+    "n_samples": 83,
+    "mse": 0.30582211903190704,
+    "rmse": 0.5530118615652896,
+    "mae": 0.297713055668107,
+    "r2": 0.06157355129088471
+  },
+  "delivery": {
+    "n_samples": 59,
+    "mse": 0.7117105624010245,
+    "rmse": 0.8436293987296937,
+    "mae": 0.5593153651077616,
+    "r2": 0.30813277436953124
+  },
+  "pdi": {
+    "n_samples": 84,
+    "accuracy": 0.6666666666666666,
+    "precision": 0.36681547619047616,
+    "recall": 0.5662442396313364,
+    "f1": 0.3804035250463822
+  },
+  "ee": {
+    "n_samples": 84,
+    "accuracy": 0.6666666666666666,
+    "precision": 0.5864396721189867,
+    "recall": 0.6201608848667672,
+    "f1": 0.5949003714961162
+  },
+  "toxic": {
+    "n_samples": 60,
+    "accuracy": 0.9333333333333333,
+    "precision": 0.7142857142857143,
+    "recall": 0.9649122807017544,
+    "f1": 0.7818181818181817
+  },
+  "biodist": {
+    "n_samples": 60,
+    "kl_divergence": 0.31884722326826903,
+    "js_divergence": 0.08589330459529161
+  }
+}

models/abl/both/20260617_232841/repeat_2/outer_fold_4/best_params.json

@@ -0,0 +1,16 @@
+{
+  "dropout": 0.41353294032230026,
+  "lr": 0.00018606616740103076,
+  "weight_decay": 9.912252717813433e-05,
+  "backbone_lr_ratio": 0.32820993966799533,
+  "moe_n_experts": 4,
+  "moe_top_k": 1,
+  "moe_expert_hidden_mult": 2,
+  "llm_lora_r": 16,
+  "d_model": 256,
+  "num_heads": 8,
+  "n_attn_layers": 4,
+  "fusion_strategy": "attention",
+  "head_hidden_dim": 128,
+  "set_transformer_block": "sab"
+}