training_scripts/training_config.json

{
  "model": "Qwen/Qwen3-4B",
  "dataset": "microsoft/rStar-Coder (synthetic_sft)",
  "samples": 10000,
  "epochs": 1,
  "lora_rank": 32,
  "lora_alpha": 64,
  "batch_size": 2,
  "gradient_accumulation": 8,
  "learning_rate": 0.0002,
  "max_seq_length": 4096,
  "results": {
    "humaneval_base": 68.9,
    "humaneval_plus": 64.0,
    "mbpp_base": 58.2,
    "mbpp_plus": 50.8
  }
}