model:
  base_model: /path/to/Baichuan2-7B-Chat

data:
  train_file: data/processed/multitask_train.jsonl

lora:
  r: 8
  alpha: 32
  dropout: 0.05
  target_modules: [q_proj, k_proj, v_proj]

train:
  output_dir: checkpoints/lora-multitask
  epochs: 3
  batch_size: 1
  gradient_accumulation_steps: 8
  learning_rate: 0.0002
  max_length: 4096

multitask:
  task_weights:
    opinion_generation: 0.45
    result_generation: 0.55
  balance_by_difficulty: true
  difficulty_range: [1, 5]