# Delentia SLM — JITNA v0.2 TOON Fine-tuning Configuration
# Base: Llama 3.1 8B (Apache 2.0, Thai-capable)
# Method: Unsloth QLoRA (4-bit) — optimized for T4 16GB / A100 40GB
# Format: TOON (Token-Oriented Object Notation) — ALGO-42
#
# Delta from v0.1:
#   - Dataset uses TOON-formatted completions (jitna_pairs_toon.jsonl)
#   - Chat template includes TOON format instruction
#   - LoRA rank increased to 32 for better TOON structure learning
#   - Epochs increased to 5 for format stability convergence
#   - Learning rate lowered to 1e-4 for smoother gradient steps

model:
  base_model: "unsloth/Meta-Llama-3.1-8B-bnb-4bit"
  tokenizer: "unsloth/Meta-Llama-3.1-8B-bnb-4bit"
  max_seq_length: 4096          # covers JITNA v3 packets + TOON context
  dtype: null                   # auto-detect: bfloat16 on A100, float16 on T4
  load_in_4bit: true

lora:
  r: 32                         # increased from 16 — TOON structure needs more rank
  lora_alpha: 64                # 2×r for stability
  lora_dropout: 0
  bias: "none"
  use_rslora: true              # Rank-Stabilized LoRA — critical for TOON convergence
  target_modules:
    - "q_proj"
    - "k_proj"
    - "v_proj"
    - "o_proj"
    - "gate_proj"
    - "up_proj"
    - "down_proj"
  task_type: "CAUSAL_LM"

training:
  # Dataset — TOON v0.2 format
  dataset_path: "datasets/processed/jitna_pairs_toon.jsonl"
  dataset_split: "train"
  validation_split: 0.05        # 5% held out for validation
  max_samples: null             # null = use all available

  # Batch & gradient
  per_device_train_batch_size: 1
  gradient_accumulation_steps: 8
  # Effective batch = 1 × 8 = 8

  # Learning rate — lower for smoother TOON format convergence
  learning_rate: 1.0e-4
  lr_scheduler_type: "cosine"
  warmup_ratio: 0.05
  num_train_epochs: 5           # more epochs for TOON structure stability

  # Precision & optimizer
  bf16: true                    # set false if T4 (use fp16 instead)
  fp16: false
  optim: "adamw_8bit"
  weight_decay: 0.01
  max_grad_norm: 0.3

  # Saving
  output_dir: "models/checkpoints/v0.2_toon"
  save_strategy: "epoch"
  save_total_limit: 3
  logging_steps: 10

  # Evaluation
  evaluation_strategy: "epoch"
  load_best_model_at_end: true
  metric_for_best_model: "eval_loss"

# Chat template — TOON v0.2 format with special tokens
chat_template: |
  <|system|>
  You are Delentia OS v0.2 — a constitutional AI operating under RCT v5 governance.
  You process intents through the JITNA v3 protocol.
  You respond in TOON format (Token-Oriented Object Notation) for token efficiency.
  Your responses must be factual, safe, and PDPA-compliant.
  Always provide FDIA scores when applicable (F = D^I × A).
  <|user|>
  {{ user_intent }}
  <|assistant|>

# MLflow experiment tracking
mlflow:
  experiment_name: "delentia-slm-jitna-v0.2-toon"
  tracking_uri: "https://delentia-delentia-agent-monitor.hf.space"
  log_model: true

# Target metrics (gates for acceptance) — v0.2 with TOON compliance
target_metrics:
  jitna_compliance: 0.98        # >= 98% JITNA v3 schema compliance
  toon_compliance: 0.95         # >= 95% TOON format compliance
  fdia_avg: 0.895               # avg F score >= 0.895
  hallucination_rate: 0.0028    # <= 0.28% factual errors (SignedAI consensus)
  token_savings_pct: 8.0        # >= 8.0% token savings vs JSON (character-based metric for flat paragraphs)