File size: 77,934 Bytes

cf4cc7a

wandb: Detected [huggingface_hub.inference, openai] in use.
wandb: Use W&B Weave for improved LLM call tracing. Install Weave with `pip install weave` then add `import weave` to the top of your script.
wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
Initial GPU Memory Stats:
  Allocated: 0.00 MB
  Cached: 0.00 MB
  Max Allocated: 0.00 MB

Initializing bidirectional cross-attention model abhinand/MedEmbed-large-v0.1...

============================================================
🦥 UNSLOTH MODE: Loading encoder with optimized kernels
============================================================

🦥 Unsloth Status:
   Available: ✅ Yes
   FastSentenceTransformer: ✅ Yes (preferred)
   Version: 2026.1.4
   CUDA: ✅ Yes
   GPU: NVIDIA GeForce RTX 4080 SUPER
   GPU Memory: 16.71 GB

   API: FastSentenceTransformer (specialized for embeddings) ✅
   Target modules: ['q_proj', 'k_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj', 'down_proj']
[INFO] Auto-detected max_seq_length for abhinand/MedEmbed-large-v0.1: 512

============================================================
🦥 UNSLOTH MODE: Creating optimized sentence encoder
============================================================
   API: FastSentenceTransformer (specialized for embeddings)
   Model: abhinand/MedEmbed-large-v0.1
   Max seq length: 512
   Full finetuning: False
Unsloth: Using fast encoder path for bert (torch.compile + SDPA)
`torch_dtype` is deprecated! Use `dtype` instead!
Unsloth: Enabled gradient checkpointing
⚠️ Unsloth loading failed: Target modules {'down_proj', 'v_proj', 'up_proj', 'k_proj', 'o_proj', 'q_proj', 'gate_proj'} not found in the base model. Please check the target modules and try again.
Traceback (most recent call last):
  File "/home/dtim/Ataur/LOKI/Cross_Attention/run_cross_attention.py", line 847, in main
    sentence_encoder = create_unsloth_sentence_encoder(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/dtim/Ataur/LOKI/Cross_Attention/unsloth_encoder.py", line 462, in create_unsloth_sentence_encoder
    sentence_encoder = FastSentenceTransformer.get_peft_model(
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/dtim/anaconda3/envs/LOKI/lib/python3.12/site-packages/unsloth/models/sentence_transformer.py", line 1685, in get_peft_model
    peft_model = peft_get_peft_model(inner_model, lora_config)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/dtim/anaconda3/envs/LOKI/lib/python3.12/site-packages/peft/mapping_func.py", line 122, in get_peft_model
    return MODEL_TYPE_TO_PEFT_MODEL_MAPPING[peft_config.task_type](
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/dtim/anaconda3/envs/LOKI/lib/python3.12/site-packages/peft/peft_model.py", line 2966, in __init__
    super().__init__(model, peft_config, adapter_name, **kwargs)
  File "/home/dtim/anaconda3/envs/LOKI/lib/python3.12/site-packages/peft/peft_model.py", line 129, in __init__
    self.base_model = cls(model, {adapter_name: peft_config}, adapter_name)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/dtim/anaconda3/envs/LOKI/lib/python3.12/site-packages/peft/tuners/tuners_utils.py", line 298, in __init__
    self.inject_adapter(self.model, adapter_name, low_cpu_mem_usage=low_cpu_mem_usage, state_dict=state_dict)
  File "/home/dtim/anaconda3/envs/LOKI/lib/python3.12/site-packages/peft/tuners/tuners_utils.py", line 880, in inject_adapter
    raise ValueError(error_msg)
ValueError: Target modules {'down_proj', 'v_proj', 'up_proj', 'k_proj', 'o_proj', 'q_proj', 'gate_proj'} not found in the base model. Please check the target modules and try again.
   Falling back to standard SentenceTransformer loading...
Encoder fine-tuning enabled (trainable_encoder=True). No encoder LoRA adapters are attached by default.
Initialized model without Flash Attention
Embedding dimension: 1024
Max sequence length: 512

Loading datasets...
Loaded 19516 training examples
⚡ Sampled 10000 training examples (deterministic, seed=42)
Loaded 2444 evaluation examples
⚡ Sampled 1000 evaluation examples (deterministic)
 Using top_k_sparse attention mechanism
 Initializing Top-K Sparse Attention (k=5)
 Initializing top-k sparse attention with method: zeros
 Successfully applied zeros initialization to top-k sparse attention
 Initializing top-k sparse attention with method: zeros
 Successfully applied zeros initialization to top-k sparse attention
Using cosine pair scoring method
Bidirectional model initialized with top_k=5, pair_score_method=cosine, share_weights=True, use_refinement=False
 Using initialization method: zeros
 Initialization parameters: {'bias_value': 0.0}
Sentence encoder is trainable
Sentence encoder dtype detected: torch.float32
Model initialized with standard cross-attention layers

📊 Model Parameter Statistics:
  Total parameters: 354,040,343
  Trainable parameters: 354,040,343
  Frozen parameters: 0
  Trainable percentage: 100.00%

🎯 Loss and Aggregation Configuration:
  Architecture: Bidirectional
  Loss type: bidirectional_triplet
  Aggregation method: top_k_pairs
  Top-k value: 5
  Norm type: rmsnorm
  Q/K RMSNorm: False
  Embedding caching: ❌ Disabled
  Triplet batching: 🔒 Isolated examples
  Triplet strategy: LIMITED
  Max triplets/example: 2

  🎲 Initialization Configuration:
    Method: zeros
    Description: Initialize all weights to zero (tests uniform attention baseline)
    Parameters: {'bias_value': 0.0}

  📊 Loss Component Weights (will be normalized):
    Triplet weight: 0.5
    Attention weight: 0.0
    Pair weight: 0.3

  ⚙️ Loss Parameters:
    Margin: 0.3
    Scale: 10.0
    Pair margin: 0.3
    Pair score method: cosine
    Share attention weights: True
    Join path extraction: True
    Join path threshold: 0.15
    Self-attention enabled: False
    Attention mechanism: top_k_sparse
    Sparse top-k: 5

  🎯 Attention Mechanism Configuration:
    Attention type: top_k_sparse
    Sparse top-k: 5
    Initialization method: zeros

  🔄 Bidirectional-Specific Configuration:
    Pair margin: 0.3
    Pair score method: cosine
    Share attention weights: True
    Join path extraction: True
    Join path threshold: 0.15
    Self-attention enabled: False
Output will be saved to: output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945

Training configuration saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_config.json

Starting Model Training...
==================================================

Starting training with ID-based triplet batches...
🎯 Training curves tracker initialized:
   📁 Output directory: output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945
   📊 Tracking batch losses: True
   📈 Tracking validation loss: True
   🔍 Tracking row-sentence metrics: True
   💾 Auto-save: True
   🎨 Auto-plot: False
🎯 Training curves tracking enabled
🏆 Test-based model saving enabled - will save best models by test metrics

🔍 Loading row-sentence evaluation data...
📊 Row-sentence evaluation data:
   Test examples: 36
   Annotations: 0
   Examples with annotations: 0
ℹ️  Row-sentence data in MIMIC format - will sync with MIMIC loader

🔍 Loading MIMIC evaluation data...
[INFO] Loaded 18 MIMIC annotations
       Diagnosis row groundings: 114
       Medication row groundings: 162
[INFO] MIMIC evaluation data loaded:
       Test examples: 36
       Annotations: 18 admissions
       Matched examples: 36
✅ MIMIC evaluation enabled (all 36 test examples)
🔄 Cache disabled: MIMIC test will be re-encoded each epoch
🔗 Syncing generic row-sentence evaluation with loaded MIMIC data for consistent AP tracking...
Preparing triplet batches...
Triplet sampling strategy: LIMITED
  Max triplets per example: 2
Using ISOLATED triplet batching (triplets from the same example stay together)
Processing examples (isolated batches): 100%|██████████| 10000/10000 [00:01<00:00, 5335.99it/s]

📊 Triplet Generation Statistics (Isolated Batching):
   Total examples: 10000
   Total triplets: 16502
   Avg triplets/example: 1.65
   Min triplets/example: 1
   Max triplets/example: 2
   Examples with 0 triplets: 0
Created 10000 triplet batches for training (all triplets included, adaptive batching)
Scheduler configured for 200000 steps with 20000 warmup steps

🔥 STAGE 0: FROZEN ENCODER ONLY EVALUATION...
Evaluating with ONLY the frozen sentence encoder (no cross-attention at all)...
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:05<00:00,  4.21s/it]
🔥 FROZEN ENCODER ONLY Accuracy: 0.806
   (Total comparisons: 24296)
🔍 Stage 0: Evaluating row-sentence alignment...
ℹ️  Using MIMIC evaluator (Level A)...
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.185
       Medication: 18 examples, F1=0.242
🔥 Stage 0 Row-Sent Avg Precision: 0.419
🔥 Stage 0 Row-Sent Overall Accuracy: 0.213

🚀 STAGE 1: SOPHISTICATED MODEL EVALUATION (PRE-TRAINING)...
Evaluating sophisticated model BEFORE training...
🚀 Using CUDA device for evaluation

🎯 TRAINING CONFIGURATION: Starting from Stage 0
   📋 Configuration detected:
      Encoder trainable: True
      Use cache: False
      Use LoRA: False
      Encoder-only training: True
   Mode: Encoder-only training (cross-attention heads frozen)
   Encoder tuning mode: gradual unfreezing
   🔓 Gradual unfreezing initialized: 2/24 top layers trainable (layers 22..23)
Enabling gradient checkpointing for memory efficiency
Optimizer groups configured: head=0 @ 1.00e-04, encoder=391 @ 1.00e-05
Initialized EncoderOnlyTripletLoss with margin=0.3, scale=10.0, ranking=infonce
Initialized EncoderOnlyTripletLoss with margin=0.3, scale=10.0
Using cosine LR schedule with warmup
🔧 Using bidirectional evaluation for bidirectional model
Evaluating bidirectional model with join path extraction...
Evaluating examples: 100%|██████████| 1000/1000 [1:10:13<00:00,  4.21s/it]
🚀 SOPHISTICATED MODEL Accuracy (Pre-training): 0.806
🔍 Stage 1: Evaluating row-sentence alignment (Sophisticated Untrained)...
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.185
       Medication: 18 examples, F1=0.242
🚀 Stage 1 Row-Sent Avg Precision: 0.418
🚀 Stage 1 Row-Sent Overall Accuracy: 0.213

📊 ARCHITECTURE BENEFIT:
   🔥→🚀 Sophisticated model benefit: +0.000 (0.0%)

📈 ROW-SENTENCE PROGRESSION (Avg Precision):
   🔥 Stage 0 (Frozen): 0.419
   🚀 Stage 1 (Sophisticated): 0.418 (+-0.000)
⚠️  WARNING: Sophisticated model barely improves over frozen encoder!
   Consider if the architecture complexity is necessary for this task.
📊 Logging initial stage metrics to wandb...
✅ Baseline metrics logged to wandb (epoch 0)

📊 Adding initial stage metrics to training curves...
📊 Initial stage metrics recorded:
   🔥 Stage 0 (Frozen): 0.81 acc, 0.42 row-sent AP
   🚀 Stage 1 (Sophisticated Untrained): 0.81 acc, 0.42 row-sent AP
   🎯 Stage 2 (Initial): 0.00 acc, 0.00 row-sent AP
📊 Adding Epoch 0 (untrained model) baseline to training curves...

🎓 STARTING TRAINING OF ENCODER ONLY (NO HEADS)
   Training model type: Encoder Only (No Heads)
   Training from Stage: 0

==================== Epoch 1/20 ====================
Training model has 26242048 trainable parameters
Epoch 1: 100%|██████████| 10000/10000 [4:48:51<00:00,  1.73s/it, batch_loss=0.687] 
Current learning rates: 5.00e-06 | margin=0.300
Epoch 1 loss statistics - Mean: 0.689, Min: 0.520, Max: 0.854
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:09:57<00:00,  4.20s/it]
Epoch 1 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36305.59it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.58it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.203
       Medication: 18 examples, F1=0.244
   Row-Sent Overall Acc: 0.224
   Row-Sent Avg Precision: 0.435
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 37328.79it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.44it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.203
       Medication: 18 examples, F1=0.244
   MIMIC Diagnosis - P: 0.119, R: 0.847, F1: 0.203
   MIMIC Medication - P: 0.153, R: 0.840, F1: 0.244

📊 Epoch 1 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 1)
   ⏱️  Epoch Time: 21544.8s
   📈 Learning Rate: 5.00e-06
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.44
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)
New best model! Accuracy: 0.806
New best test overall accuracy! 0.224 (Epoch 1)
New best test average precision! 0.435 (Epoch 1)

==================== Epoch 2/20 ====================
   🔓 Gradual unfreezing: enabled layers 21..21 (total 3/24)
Epoch 2: 100%|██████████| 10000/10000 [4:49:35<00:00,  1.74s/it, batch_loss=0.683] 
Current learning rates: 1.00e-05 | margin=0.300
Epoch 2 loss statistics - Mean: 0.689, Min: 0.523, Max: 0.871
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:21<00:00,  4.22s/it]
Epoch 2 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36314.32it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.60it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.203
       Medication: 18 examples, F1=0.244
   Row-Sent Overall Acc: 0.223
   Row-Sent Avg Precision: 0.435
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 37748.74it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.77it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.17it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.203
       Medication: 18 examples, F1=0.244
   MIMIC Diagnosis - P: 0.118, R: 0.847, F1: 0.203
   MIMIC Medication - P: 0.153, R: 0.840, F1: 0.244

📊 Epoch 2 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 1)
   ⏱️  Epoch Time: 21613.4s
   📈 Learning Rate: 1.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.44
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 3/20 ====================
   🔓 Gradual unfreezing: enabled layers 20..20 (total 4/24)
Epoch 3: 100%|██████████| 10000/10000 [4:50:32<00:00,  1.74s/it, batch_loss=0.683] 
Current learning rates: 9.92e-06 | margin=0.300
Epoch 3 loss statistics - Mean: 0.689, Min: 0.504, Max: 0.875
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:17<00:00,  4.22s/it]
Epoch 3 Accuracy: 0.805
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 35511.51it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.69it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.243
   Row-Sent Overall Acc: 0.222
   Row-Sent Avg Precision: 0.435
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36846.01it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.21it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.17it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.244
   MIMIC Diagnosis - P: 0.117, R: 0.847, F1: 0.201
   MIMIC Medication - P: 0.153, R: 0.840, F1: 0.244

📊 Epoch 3 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 1)
   ⏱️  Epoch Time: 21666.8s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.44
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 4/20 ====================
   🔓 Gradual unfreezing: enabled layers 19..19 (total 5/24)
Epoch 4: 100%|██████████| 10000/10000 [4:50:41<00:00,  1.74s/it, batch_loss=0.683] 
Current learning rates: 9.70e-06 | margin=0.300
Epoch 4 loss statistics - Mean: 0.689, Min: 0.525, Max: 0.854
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:11<00:00,  4.21s/it]
Epoch 4 Accuracy: 0.805
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36756.32it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.76it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.17it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.244
   Row-Sent Overall Acc: 0.222
   Row-Sent Avg Precision: 0.435
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 18810.88it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 56.02it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.244
   MIMIC Diagnosis - P: 0.117, R: 0.847, F1: 0.201
   MIMIC Medication - P: 0.153, R: 0.840, F1: 0.244

📊 Epoch 4 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 1)
   ⏱️  Epoch Time: 21669.1s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 5/20 ====================
   🔓 Gradual unfreezing: enabled layers 18..18 (total 6/24)
Epoch 5: 100%|██████████| 10000/10000 [4:50:47<00:00,  1.74s/it, batch_loss=0.686] 
Current learning rates: 9.33e-06 | margin=0.300
Epoch 5 loss statistics - Mean: 0.689, Min: 0.535, Max: 0.854
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:09:56<00:00,  4.20s/it]
Epoch 5 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36054.19it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.85it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.243
   Row-Sent Overall Acc: 0.222
   Row-Sent Avg Precision: 0.434
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 39097.60it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.19it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.243
   MIMIC Diagnosis - P: 0.117, R: 0.849, F1: 0.201
   MIMIC Medication - P: 0.153, R: 0.840, F1: 0.243

📊 Epoch 5 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 1)
   ⏱️  Epoch Time: 21660.5s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 6/20 ====================
Epoch 6: 100%|██████████| 10000/10000 [4:50:51<00:00,  1.75s/it, batch_loss=0.688] 
Current learning rates: 8.83e-06 | margin=0.300
Epoch 6 loss statistics - Mean: 0.689, Min: 0.532, Max: 0.857
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:18<00:00,  4.22s/it]
Epoch 6 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 35603.62it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.66it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.244
   Row-Sent Overall Acc: 0.222
   Row-Sent Avg Precision: 0.434
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36756.32it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 58.97it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.17it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.244
   MIMIC Diagnosis - P: 0.117, R: 0.851, F1: 0.201
   MIMIC Medication - P: 0.153, R: 0.840, F1: 0.244

📊 Epoch 6 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 1)
   ⏱️  Epoch Time: 21686.2s
   📈 Learning Rate: 8.83e-06
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 7/20 ====================
Epoch 7: 100%|██████████| 10000/10000 [4:51:28<00:00,  1.75s/it, batch_loss=0.682] 
Current learning rates: 8.21e-06 | margin=0.300
Epoch 7 loss statistics - Mean: 0.689, Min: 0.512, Max: 0.866
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:17<00:00,  4.22s/it]
Epoch 7 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 37504.95it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.92it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.16it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.243
   Row-Sent Overall Acc: 0.222
   Row-Sent Avg Precision: 0.435
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 16710.37it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 57.66it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.243
   MIMIC Diagnosis - P: 0.117, R: 0.851, F1: 0.201
   MIMIC Medication - P: 0.152, R: 0.840, F1: 0.243

📊 Epoch 7 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 1)
   ⏱️  Epoch Time: 21722.1s
   📈 Learning Rate: 8.21e-06
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 8/20 ====================
Epoch 8: 100%|██████████| 10000/10000 [4:51:15<00:00,  1.75s/it, batch_loss=0.682] 
Current learning rates: 7.50e-06 | margin=0.300
Epoch 8 loss statistics - Mean: 0.689, Min: 0.536, Max: 0.865
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:14<00:00,  4.21s/it]
Epoch 8 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 37099.49it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.28it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.201
       Medication: 18 examples, F1=0.243
   Row-Sent Overall Acc: 0.222
   Row-Sent Avg Precision: 0.434
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36587.10it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.93it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.200
       Medication: 18 examples, F1=0.243
   MIMIC Diagnosis - P: 0.117, R: 0.851, F1: 0.200
   MIMIC Medication - P: 0.152, R: 0.840, F1: 0.243

📊 Epoch 8 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 1)
   ⏱️  Epoch Time: 21706.6s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 9/20 ====================
Epoch 9: 100%|██████████| 10000/10000 [4:51:10<00:00,  1.75s/it, batch_loss=0.681] 
Current learning rates: 6.71e-06 | margin=0.300
Epoch 9 loss statistics - Mean: 0.689, Min: 0.522, Max: 0.859
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:14<00:00,  4.21s/it]
Epoch 9 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 33149.27it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.62it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.200
       Medication: 18 examples, F1=0.241
   Row-Sent Overall Acc: 0.221
   Row-Sent Avg Precision: 0.434
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 35704.65it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 58.34it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.200
       Medication: 18 examples, F1=0.241
   MIMIC Diagnosis - P: 0.116, R: 0.851, F1: 0.200
   MIMIC Medication - P: 0.151, R: 0.840, F1: 0.241

📊 Epoch 9 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 1)
   ⏱️  Epoch Time: 21701.2s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 10/20 ====================
Epoch 10: 100%|██████████| 10000/10000 [4:51:09<00:00,  1.75s/it, batch_loss=0.678] 
Current learning rates: 5.87e-06 | margin=0.300
Epoch 10 loss statistics - Mean: 0.689, Min: 0.531, Max: 0.860
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:17<00:00,  4.22s/it]
Epoch 10 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 34735.44it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.24it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.200
       Medication: 18 examples, F1=0.241
   Row-Sent Overall Acc: 0.221
   Row-Sent Avg Precision: 0.434
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 38101.17it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.52it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.200
       Medication: 18 examples, F1=0.241
   MIMIC Diagnosis - P: 0.116, R: 0.851, F1: 0.200
   MIMIC Medication - P: 0.151, R: 0.840, F1: 0.241

📊 Epoch 10 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21704.0s
   📈 Learning Rate: 5.87e-06
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)
New best model! Accuracy: 0.806

==================== Epoch 11/20 ====================
Epoch 11: 100%|██████████| 10000/10000 [4:51:28<00:00,  1.75s/it, batch_loss=0.678] 
Current learning rates: 5.00e-06 | margin=0.300
Epoch 11 loss statistics - Mean: 0.689, Min: 0.530, Max: 0.859
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:21<00:00,  4.22s/it]
Epoch 11 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 35229.80it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.53it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.17it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.200
       Medication: 18 examples, F1=0.241
   Row-Sent Overall Acc: 0.221
   Row-Sent Avg Precision: 0.433
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 34976.82it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.47it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.17it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.200
       Medication: 18 examples, F1=0.241
   MIMIC Diagnosis - P: 0.116, R: 0.851, F1: 0.200
   MIMIC Medication - P: 0.151, R: 0.840, F1: 0.241

📊 Epoch 11 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21726.3s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 12/20 ====================
Epoch 12: 100%|██████████| 10000/10000 [4:51:23<00:00,  1.75s/it, batch_loss=0.685] 
Current learning rates: 4.13e-06 | margin=0.300
Epoch 12 loss statistics - Mean: 0.689, Min: 0.508, Max: 0.865
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:16<00:00,  4.22s/it]
Epoch 12 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 33391.19it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.82it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.200
       Medication: 18 examples, F1=0.240
   Row-Sent Overall Acc: 0.220
   Row-Sent Avg Precision: 0.433
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 37588.98it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.78it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.11it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.200
       Medication: 18 examples, F1=0.240
   MIMIC Diagnosis - P: 0.116, R: 0.847, F1: 0.200
   MIMIC Medication - P: 0.149, R: 0.840, F1: 0.240

📊 Epoch 12 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21717.2s
   📈 Learning Rate: 4.13e-06
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 13/20 ====================
Epoch 13: 100%|██████████| 10000/10000 [4:51:05<00:00,  1.75s/it, batch_loss=0.684] 
Current learning rates: 3.29e-06 | margin=0.300
Epoch 13 loss statistics - Mean: 0.689, Min: 0.528, Max: 0.890
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:09:57<00:00,  4.20s/it]
Epoch 13 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36711.63it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.63it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   Row-Sent Overall Acc: 0.219
   Row-Sent Avg Precision: 0.433
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 39808.84it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.23it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   MIMIC Diagnosis - P: 0.115, R: 0.847, F1: 0.198
   MIMIC Medication - P: 0.149, R: 0.840, F1: 0.240

📊 Epoch 13 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21679.7s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 14/20 ====================
Epoch 14: 100%|██████████| 10000/10000 [4:50:45<00:00,  1.74s/it, batch_loss=0.686] 
Current learning rates: 2.50e-06 | margin=0.300
Epoch 14 loss statistics - Mean: 0.689, Min: 0.542, Max: 0.859
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:09:57<00:00,  4.20s/it]
Epoch 14 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36114.55it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.32it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   Row-Sent Overall Acc: 0.219
   Row-Sent Avg Precision: 0.433
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 38796.23it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.26it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   MIMIC Diagnosis - P: 0.115, R: 0.847, F1: 0.198
   MIMIC Medication - P: 0.149, R: 0.840, F1: 0.240

📊 Epoch 14 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21660.1s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 15/20 ====================
Epoch 15: 100%|██████████| 10000/10000 [4:50:47<00:00,  1.74s/it, batch_loss=0.688] 
Current learning rates: 1.79e-06 | margin=0.300
Epoch 15 loss statistics - Mean: 0.689, Min: 0.532, Max: 0.869
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:03<00:00,  4.20s/it]
Epoch 15 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 37356.49it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.65it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   Row-Sent Overall Acc: 0.219
   Row-Sent Avg Precision: 0.433
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36428.21it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.69it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   MIMIC Diagnosis - P: 0.115, R: 0.847, F1: 0.198
   MIMIC Medication - P: 0.149, R: 0.840, F1: 0.240

📊 Epoch 15 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21666.9s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 16/20 ====================
Epoch 16: 100%|██████████| 10000/10000 [4:50:44<00:00,  1.74s/it, batch_loss=0.689] 
Current learning rates: 1.17e-06 | margin=0.300
Epoch 16 loss statistics - Mean: 0.689, Min: 0.534, Max: 0.860
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:09:57<00:00,  4.20s/it]
Epoch 16 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 36393.09it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.83it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   Row-Sent Overall Acc: 0.219
   Row-Sent Avg Precision: 0.433
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 39250.05it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.32it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   MIMIC Diagnosis - P: 0.115, R: 0.847, F1: 0.198
   MIMIC Medication - P: 0.149, R: 0.840, F1: 0.240

📊 Epoch 16 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21658.9s
   📈 Learning Rate: 1.17e-06
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 17/20 ====================
Epoch 17: 100%|██████████| 10000/10000 [4:50:45<00:00,  1.74s/it, batch_loss=0.686] 
Current learning rates: 1.00e-06 | margin=0.300
Epoch 17 loss statistics - Mean: 0.689, Min: 0.534, Max: 0.849
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:09:57<00:00,  4.20s/it]
Epoch 17 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 37467.73it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.80it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   Row-Sent Overall Acc: 0.219
   Row-Sent Avg Precision: 0.433
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 37035.80it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.21it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   MIMIC Diagnosis - P: 0.115, R: 0.847, F1: 0.198
   MIMIC Medication - P: 0.149, R: 0.840, F1: 0.240

📊 Epoch 17 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21660.2s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 18/20 ====================
Epoch 18: 100%|██████████| 10000/10000 [4:50:43<00:00,  1.74s/it, batch_loss=0.679] 
Current learning rates: 1.00e-06 | margin=0.300
Epoch 18 loss statistics - Mean: 0.689, Min: 0.532, Max: 0.872
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:09:58<00:00,  4.20s/it]
Epoch 18 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 37090.38it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.83it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   Row-Sent Overall Acc: 0.219
   Row-Sent Avg Precision: 0.433
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 34647.76it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.39it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.19it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   MIMIC Diagnosis - P: 0.115, R: 0.847, F1: 0.198
   MIMIC Medication - P: 0.149, R: 0.840, F1: 0.240

📊 Epoch 18 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21658.0s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 19/20 ====================
Epoch 19: 100%|██████████| 10000/10000 [4:50:59<00:00,  1.75s/it, batch_loss=0.687] 
Current learning rates: 1.00e-06 | margin=0.300
Epoch 19 loss statistics - Mean: 0.689, Min: 0.518, Max: 0.866
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:17<00:00,  4.22s/it]
Epoch 19 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 32640.50it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.64it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   Row-Sent Overall Acc: 0.219
   Row-Sent Avg Precision: 0.433
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 34799.48it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 52.03it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   MIMIC Diagnosis - P: 0.115, R: 0.847, F1: 0.198
   MIMIC Medication - P: 0.149, R: 0.840, F1: 0.240

📊 Epoch 19 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21693.0s
   📈 Learning Rate: 1.00e-06
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

==================== Epoch 20/20 ====================
Epoch 20: 100%|██████████| 10000/10000 [4:51:12<00:00,  1.75s/it, batch_loss=0.684] 
Current learning rates: 1.00e-06 | margin=0.300
Epoch 20 loss statistics - Mean: 0.689, Min: 0.538, Max: 0.871
Loss not decreasing. Patience: 1/1
Reset learning rate to 5.00e-05
Evaluating after epoch...
🔎 Validation evaluator: Stage 0 (frozen encoder baseline)
🔥 Evaluating with FROZEN ENCODER ONLY (no cross-attention, aggregation=top_k_pairs)
Evaluating frozen encoder baseline: 100%|██████████| 1000/1000 [1:10:14<00:00,  4.21s/it]
Epoch 20 Accuracy: 0.806
Collecting tables and contexts for test_row_sent_epoch split...
Processing test_row_sent_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 34560.53it/s]
Found 36 unique tables and 39 unique contexts in test_row_sent_epoch split
Using device: cuda for embedding cache
Encoding test_row_sent_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 61.65it/s]
Encoding test_row_sent_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.18it/s]
🔎 Row-sent evaluator: Stage 0 (frozen encoder baseline)
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   Row-Sent Overall Acc: 0.219
   Row-Sent Avg Precision: 0.433
   Examples evaluated: 36
Collecting tables and contexts for mimic_test_epoch split...
Processing mimic_test_epoch examples: 100%|██████████| 36/36 [00:00<00:00, 38333.32it/s]
Found 36 unique tables and 39 unique contexts in mimic_test_epoch split
Using device: cuda for embedding cache
Encoding mimic_test_epoch tables: 100%|██████████| 36/36 [00:00<00:00, 59.05it/s]
Encoding mimic_test_epoch contexts: 100%|██████████| 39/39 [00:07<00:00,  5.14it/s]
[INFO] MIMIC Row Grounding evaluation: 36 examples
       Diagnosis: 18 examples, F1=0.198
       Medication: 18 examples, F1=0.240
   MIMIC Diagnosis - P: 0.115, R: 0.847, F1: 0.198
   MIMIC Medication - P: 0.149, R: 0.840, F1: 0.240

📊 Epoch 20 Summary:
   🔥 Train Loss: 0.69 ± 0.04
   🎯 Val Accuracy: 0.81
   🏆 Best Accuracy: 0.81 (Epoch 10)
   ⏱️  Epoch Time: 21702.7s
   📈 Learning Rate: 5.00e-05
   🔍 Row-Sent Overall Acc: 0.22
      🏆 Best Test Overall Acc: 0.22 (Epoch 1)
   🔍 Row-Sent Avg Precision: 0.43
      🏆 Best Test Avg Precision: 0.44 (Epoch 1)
   Stage 0 note: encoder fine-tuning active; eval uses on-the-fly embeddings (no cache)

Loading best model from epoch 10 with accuracy 0.806
Best model saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/abhinand/MedEmbed-large-v0.1/best_model_epoch_10/model.pt

🎯 SAVING BEST TEST-BASED MODELS:
Loading best test overall accuracy model from epoch 1 (0.224)
Best test overall accuracy model saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/abhinand/MedEmbed-large-v0.1/best_test_overall_acc_epoch_1/model.pt
Loading best test average precision model from epoch 1 (0.435)
Best test average precision model saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/abhinand/MedEmbed-large-v0.1/best_test_avg_precision_epoch_1/model.pt
Reloaded validation-based best model for final analysis

==========================================================================================
🎯 COMPREHENSIVE 3-STAGE IMPACT ANALYSIS
==========================================================================================
📊 COMPLETE PERFORMANCE BREAKDOWN:
   🔥 Stage 0 - Frozen Encoder Only:     0.806
   🚀 Stage 1 - Sophisticated (Pre):     0.806 (+0.000)
   🏆 Stage 2 - Trained Model:           0.806 (+0.000)
   📈 Total Improvement:                 +0.000 (0.0%)

🏆 TEST METRICS SUMMARY:
   🎯 Best Test Overall Accuracy:     0.224 (Epoch 1)
   🎯 Best Test Average Precision:    0.435 (Epoch 1)
   📊 Validation vs Test Performance:
      - Validation Best: 0.806 (Epoch 10)
      - Test Overall Acc: 0.224 (Epoch 1)
      - Test Avg Precision: 0.435 (Epoch 1)
   ⚠️  NOTE: Best test overall accuracy occurred at epoch 1, not at best validation epoch 10✅ Final summary logged to wandb

🎨 Skipping 3-STAGE example visualizations (skip_four_stage_viz=True)
==========================================================================================

======================================================================
🎯 TRAINING CURVES SUMMARY
======================================================================

======================================================================
🎯 TRAINING SUMMARY - abhinand/MedEmbed-large-v0.1
======================================================================
📊 Total Epochs: 20
🏆 Best Accuracy: 0.81 (Epoch 10)
📈 Final Accuracy: 0.81
📉 Accuracy Improvement: -0.00
🔥 Initial Loss: 0.69
🎯 Final Loss: 0.69
📉 Loss Reduction: -0.00
⏱️  Average Epoch Time: 21674.9s
🕐 Total Training Time: 433497.5s (7225.0 min)
🏆 Best Test Metrics:
   🔍 Best Overall Accuracy: 0.22 (Epoch 1)
   🔍 Best Average Precision: 0.44 (Epoch 1)
📊 Initial Stage Metrics:
   🔥 Stage 0 (Frozen): 0.81 acc, 0.42 row-sent AP
   🚀 Stage 1 (Sophisticated Untrained): 0.81 acc, 0.42 row-sent AP
   🎯 Stage 2 (Initial): 0.00 acc, 0.00 row-sent AP
======================================================================
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_final_training_curves.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_final_training_curves.pdf
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_training_loss.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_training_loss.pdf
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_validation_accuracy.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_validation_accuracy.pdf
📈 Generating batch-level analysis...
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_batch_losses_heatmap.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_batch_losses_heatmap.pdf
🌡️ Batch losses heatmap saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_batch_losses_heatmap.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_batch_losses_epoch_1.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_batch_losses_epoch_1.pdf
📈 Batch losses plot for epoch 1 saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_batch_losses_epoch_1.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_batch_losses_epoch_21.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_batch_losses_epoch_21.pdf
📈 Batch losses plot for epoch 21 saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots/abhinand_MedEmbed-large-v0.1_batch_losses_epoch_21.png
✅ Training curves analysis complete!
📁 All plots saved to: output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_plots
💾 Training data saved to: output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/training_data

🎓 TRAINING COMPLETED: Returning trained Encoder Only (No Heads)

✅ Training completed! The returned model is the BEST performing model from training.
📁 Check the training logs above to see which epoch achieved the highest validation accuracy.
Saved best model to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/MedEmbed-large-v0.1_best.pt

Evaluating on Test Set...
Loaded 36 test examples
Cache disabled - evaluation will compute embeddings on-the-fly
Evaluating bidirectional model with join path extraction...
Evaluating examples: 100%|██████████| 36/36 [00:18<00:00,  1.94it/s]
Saved 36 join path examples to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/extracted_join_paths.json
Test Accuracy: 0.931
Join Paths Extracted: 36
Evaluation results saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/test_metrics.json

🎨 Generating visualizations using the BEST model (highest validation accuracy)...
✅ CONFIRMED: Using BEST model checkpoint from training
📁 Best model path: output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/MedEmbed-large-v0.1_best.pt
🔧 Using aggregation method: top_k_pairs
✅ Best model file confirmed to exist

Running comprehensive visualizations using top_k_pairs...
🔍 Loading trained model from: output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/MedEmbed-large-v0.1_best.pt
📁 Loading TRAINED model from checkpoint: output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/MedEmbed-large-v0.1_best.pt
✅ CONFIRMED: Loading BEST model checkpoint (highest validation accuracy)
📄 Found configuration file: args.json
✅ Loaded from config: model_type=Bidirectional, embedding_dim=1024, attention_type=top_k_sparse
Creating dynamic encoder for 1024 dimensions on cuda
📐 Final embedding dimension: 1024
Creating BidirectionalTableTextModel...
🚀 Using top_k_sparse attention mechanism
📋 Using top-5 sparse attention
 Using top_k_sparse attention mechanism
 Initializing Top-K Sparse Attention (k=5)
 Initializing top-k sparse attention with method: xavier_uniform
 Successfully applied xavier_uniform initialization to top-k sparse attention
 Initializing top-k sparse attention with method: xavier_uniform
 Successfully applied xavier_uniform initialization to top-k sparse attention
Using cosine pair scoring method
Bidirectional model initialized with top_k=3, pair_score_method=cosine, share_weights=False, use_refinement=True
 Using initialization method: xavier_uniform
Sentence encoder is frozen
Sentence encoder dtype detected: torch.float32
Loading trained weights from output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/MedEmbed-large-v0.1_best.pt
✅ Trained weights loaded successfully (custom layers only)
🏆 VERIFICATION: Best model weights are now active for visualization
🎯 Model ready for visualization: BEST TRAINED model on cuda
✅ Trained model loaded with dynamic dimension detection on cuda

Processing example 16 (ID: 2221973625275923611):
  8 rows, 69 sentences
  Using aggregation method: top_k_pairs

Generating comprehensive 4-panel analysis...

🔍 Creating comprehensive analysis for Example 16...
  Found 8 rows and 69 sentences
Computing comprehensive similarities...
  Computing raw embedding similarities...
  Computing bidirectional attention and contextualized similarities...
 Applying top_k_sparse attention mechanism...
  Computing contextualized similarities...
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/comprehensive_analysis_example_16.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/comprehensive_analysis_example_16.pdf
💾 Saved comprehensive analysis to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/comprehensive_analysis_example_16.png

📊 Analysis Summary for Example 16:
  Raw Embeddings - Range: [-0.101, 0.082]
  Contextualized Similarities - Range: [-0.071, 0.396]
  Final Model Similarities - Range: [-0.091, 0.082]
  Cross-Attention - Range: [0.000, 0.200]

🔝 Top 3 pairs by Raw Embeddings:
  1. Row 2 - Sentence 57: 0.082
  2. Row 6 - Sentence 55: 0.079
  3. Row 6 - Sentence 45: 0.074

🔝 Top 3 pairs by Contextualized Similarities:
  1. Row 1 - Sentence 3: 0.396
  2. Row 3 - Sentence 2: 0.379
  3. Row 5 - Sentence 4: 0.377

🔝 Top 3 pairs by Final Model Similarities:
  1. Row 4 - Sentence 27: 0.082
  2. Row 5 - Sentence 20: 0.082
  3. Row 4 - Sentence 10: 0.079

🔝 Top 3 pairs by Cross-Attention:
  1. Row 5 - Sentence 2: 0.200
  2. Row 5 - Sentence 3: 0.200
  3. Row 5 - Sentence 4: 0.200
📄 Detailed analysis saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/example_16_similarity_analysis.txt


All visualizations saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations

🔬 Creating comprehensive model analysis...
🔬 Creating comprehensive analysis for example 16 using BEST model...

🔍 Creating comprehensive analysis for Example 16...
  Found 8 rows and 69 sentences
Computing comprehensive similarities...
  Computing raw embedding similarities...
  Computing bidirectional attention and contextualized similarities...
 Applying top_k_sparse attention mechanism...
  Computing contextualized similarities...
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/comprehensive_analysis_example_16.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/comprehensive_analysis_example_16.pdf
💾 Saved comprehensive analysis to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/comprehensive_analysis_example_16.png

📊 Analysis Summary for Example 16:
  Raw Embeddings - Range: [0.456, 0.725]
  Contextualized Similarities - Range: [0.834, 0.949]
  Final Model Similarities - Range: [0.456, 0.725]
  Cross-Attention - Range: [0.000, 0.200]

🔝 Top 3 pairs by Raw Embeddings:
  1. Row 8 - Sentence 25: 0.725
  2. Row 6 - Sentence 25: 0.723
  3. Row 2 - Sentence 28: 0.717

🔝 Top 3 pairs by Contextualized Similarities:
  1. Row 4 - Sentence 38: 0.949
  2. Row 1 - Sentence 2: 0.943
  3. Row 4 - Sentence 4: 0.943

🔝 Top 3 pairs by Final Model Similarities:
  1. Row 8 - Sentence 25: 0.725
  2. Row 6 - Sentence 25: 0.723
  3. Row 2 - Sentence 28: 0.717

🔝 Top 3 pairs by Cross-Attention:
  1. Row 5 - Sentence 2: 0.200
  2. Row 5 - Sentence 3: 0.200
  3. Row 5 - Sentence 4: 0.200
📄 Detailed analysis saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/example_16_similarity_analysis.txt

🔬 Generating step-by-step diagnostics for example 16...
🔬 Generating step-by-step diagnostics for example 16...
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step1_raw_similarities.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step1_raw_similarities.pdf
 Applying top_k_sparse attention mechanism...
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step2_forward_attention.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step2_forward_attention.pdf
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step3_reverse_attention.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step3_reverse_attention.pdf
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step4_contextualized_similarities.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step4_contextualized_similarities.pdf
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step5_refined_similarities.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step5_refined_similarities.pdf
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step6_final_pair_scores.png
Saved plot to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/step6_final_pair_scores.pdf
  Generating validation report...
  Validation report saved to output_best_model/bidirectional_MedEmbed-large-v0.1_20260128_011945/visualizations/diagnostics_example_16/validation_report_example_16.txt

🎯 Skipping 4-STAGE example visualizations (--skip_four_stage_viz is set)