model:
  name: "Qwen/Qwen3-4B-Instruct-2507"
  torch_dtype: "bfloat16"
  trust_remote_code: false

lora:
  r: 16
  lora_alpha: 32
  target_modules: ["q_proj", "v_proj", "k_proj", "o_proj"]
  lora_dropout: 0.05
  bias: "none"

training:
  num_epochs: 3
  per_device_batch_size: 2
  gradient_accumulation_steps: 3
  learning_rate: 0.0002
  lr_scheduler: "cosine"
  warmup_ratio: 0.05
  bf16: true
  gradient_checkpointing: false
  logging_steps: 20
  save_strategy: "epoch"
  save_total_limit: 3
  max_length: 1024

generation:
  max_new_tokens: 512
  temperature: 1.0
  n_samples: 32

paths:
  checkpoint_dir: "checkpoints/"
  results_dir: "results/"
  data_dir: "data/"

lean_server:
  url: "http://localhost:8000"
  timeout: 60
  max_workers: 32