{
  "base_model": "Qwen/Qwen3-4B",
  "base_model_revision": "1cfa9a7208912126459214e8b04321603b3df60c",
  "task": "math-hard",
  "method": "supervised fine-tuning (LoRA)",
  "enable_thinking": false,
  "training_conversations": 450,
  "validation_conversations": 50,
  "global_step": 24,
  "epochs_completed": 3.0,
  "data_source": "https://github.com/Wuyxin/collabllm/tree/d573e4949abdc301cb2b630525c593cc432e94b8/data/interactive-math-hard",
  "data_preparation": "For each dialogue idx, select the row with the longest prompt and append its chosen assistant response. Hold out 50 of 500 conversations using seed 42.",
  "adapter_sha256": "ac1e634d0ab09341a38a4be7e54959edccf94509fffd6d448d7ad7f52c1b0e9e",
  "hyperparameters": {
    "seed": 42,
    "learning_rate": 1e-05,
    "warmup_ratio": 0.1,
    "max_seq_length": 1536,
    "num_train_epochs": 3,
    "per_device_train_batch_size": 1,
    "gradient_accumulation_steps": 64,
    "assistant_only_loss": true,
    "peft_r": 32,
    "peft_alpha": 16,
    "peft_dropout": 0.1,
    "target_modules": "q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj",
    "use_4bit": false
  },
  "training_metrics": {
    "total_flos": 4.277110198134374e+16,
    "train_loss": 0.8463213642438253,
    "train_runtime": 1324.9645,
    "train_samples_per_second": 1.019,
    "train_steps_per_second": 0.018
  },
  "validation_metrics": {
    "eval_loss": 0.7929708361625671,
    "eval_runtime": 13.7847,
    "eval_samples_per_second": 3.627,
    "eval_steps_per_second": 3.627
  }
}