training_metadata.json

{
  "model_info": {
    "name": "conservative_medical_phase2_20250724_024052",
    "type": "conservative_few_shot_prototypical",
    "phase": 2,
    "base_model": "emilyalsentzer/Bio_ClinicalBERT"
  },
  "performance": {
    "top_1_accuracy": 0.463,
    "top_3_accuracy": "~0.52",
    "top_5_accuracy": "~0.54",
    "baseline_improvement": "+6.8pp",
    "baseline_accuracy": 0.395
  },
  "architecture": {
    "embedding_dim": 768,
    "pooling_strategy": "cls_plus_mean",
    "prototype_method": "ensemble_attention",
    "loss_function": "simple_triplet"
  },
  "optimizations": {
    "phase_1": "larger_embeddings_enhanced_pooling",
    "phase_2": "ensemble_prototypes_attention_aggregation",
    "conservative_approach": true,
    "stable_training": true
  },
  "deployment": {
    "ready": true,
    "requires_dataset": false,
    "self_contained": true,
    "num_medical_codes": "~57000"
  },
  "usage": {
    "input": "clinical_notes_text",
    "output": "top_k_medical_codes",
    "recommended_top_k": [
      1,
      3,
      5,
      10
    ],
    "inference_method": "prototype_similarity"
  },
  "training": {
    "epochs": 3,
    "stable_loss": true,
    "no_collapse": true,
    "incremental_improvement": true
  },
  "timestamp": "20250724_024052",
  "version": "2.0"
}