# SFT Training Config - Local Data, Multi-GPU
# Usage: accelerate launch --config_file configs/accelerate_multi_gpu.yaml train.py sft --config configs/sft_local.yaml

model:
  base: "Qwen/Qwen3-1.7B"
  output: "outputs/sft"  # Local output
  push_to_hub: false

dataset:
  name: "data/train"  # Local path
  text_field: "text"
  split: "train"
  eval_split: 0.1

training:
  epochs: 5
  batch_size: 2  # Per GPU, effective batch = 2 * 4 GPUs * 4 accum = 32
  gradient_accumulation_steps: 4
  learning_rate: 0.0002  # 2e-4 as float
  max_length: 512
  warmup_ratio: 0.03
  lr_scheduler: "cosine"
  ddp_find_unused_parameters: false
  # Save checkpoints every 30 minutes
  save_interval_minutes: 30
  # Fallback time-step save cadence if needed (not used for wall-clock mode)
  save_steps: 200

lora:
  rank: 16
  alpha: 32
  dropout: 0.05
  target_modules:
    - "q_proj"
    - "k_proj"
    - "v_proj"
    - "o_proj"
    - "gate_proj"
    - "up_proj"
    - "down_proj"

tracking:
  project: "qmd-query-expansion"
  run_name: "{day} {time}"