# DCE Intent BioClinicalBERT Training Configuration
# ============================================
# Version: 2.0.0
# Last Updated: 2026-02-08

# Model Configuration
# Using BioClinicalBERT for medical domain accuracy
# Reduced max_length to 48 to maintain low latency
model:
  base_model: "emilyalsentzer/Bio_ClinicalBERT"
  max_length: 48  # Reduced from 64 to compensate for larger model
  num_labels: null  # Auto-detected from dataset

# Training Hyperparameters
training:
  learning_rate: 3.0e-5
  epochs: 15
  batch_size: 16
  eval_batch_size: 32
  warmup_ratio: 0.1
  weight_decay: 0.01
  gradient_accumulation_steps: 1
  max_grad_norm: 1.0
  
# Optimizer
optimizer:
  type: "adamw"
  betas: [0.9, 0.999]
  eps: 1.0e-8

# Scheduler
scheduler:
  type: "linear"  # linear warmup then linear decay

# Data Paths
data:
  train_path: "data/examples/train.jsonl"
  val_path: "data/examples/val.jsonl"
  test_path: "data/examples/test.jsonl"
  text_column: "text"
  label_column: "label"

# Output Configuration
output:
  base_dir: "artifacts"
  pytorch_dir: "artifacts/pytorch"
  save_steps: 30
  eval_steps: 15
  logging_steps: 15
  save_total_limit: 2

# Reproducibility
seed: 42
deterministic: true

# Logging
logging:
  level: "INFO"
  log_to_file: true
  log_file: "artifacts/logs/training.log"

# Intent Schema (Tier 1 - Required)
intent_schema:
  tier1:
    - "ESCALATION"          # Urgent medical escalation
    - "APPOINTMENT"         # Schedule/reschedule appointments
    - "MEDICATION"          # Medication-related queries
    - "SYMPTOM_CHECK"       # Symptom assessment
    - "GENERAL_INQUIRY"     # General health questions
    - "BILLING"             # Billing/insurance questions
    - "OTHER"               # Fallback category
  tier2:  # Optional sub-categories
    ESCALATION:
      - "CHEST_PAIN"
      - "BREATHING_DIFFICULTY"
      - "SEVERE_BLEEDING"
      - "LOSS_OF_CONSCIOUSNESS"
      - "OTHER_EMERGENCY"

# Evaluation Metrics
metrics:
  primary: "macro_f1"
  track:
    - "accuracy"
    - "macro_f1"
    - "per_class_recall"
  escalation_label: "ESCALATION"  # Label to track recall for