training_args.json

{
  "output_dir": "/mnt/models/runs/sft-qwen2.5-7b-k8s-001",
  "num_train_epochs": 3,
  "max_steps": 1200,
  "per_device_train_batch_size": 1,
  "gradient_accumulation_steps": 16,
  "learning_rate": 0.0002,
  "lr_scheduler_type": "cosine",
  "warmup_ratio": 0.03,
  "bf16": true,
  "gradient_checkpointing": true,
  "evaluation_strategy": "steps",
  "eval_steps": 100,
  "save_steps": 120,
  "save_total_limit": 3,
  "seed": 42
}