training_args.json

{
  "model_name": "Qwen/Qwen2.5-3B",
  "dataset_name": "nvidia/OpenMathInstruct-2",
  "max_length": 2048,
  "max_samples": 100000,
  "batch_size": 3,
  "gradient_accumulation_steps": 16,
  "learning_rate": 2e-05,
  "num_epochs": 3,
  "warmup_steps": 200,
  "weight_decay": 0.01,
  "seed": 42,
  "max_checkpoints": 3,
  "save_steps": 10000,
  "eval_steps": 10000,
  "output_dir": "./qwen_math_fp8_model",
  "fp8_backend": "msamp",
  "msamp_opt_level": "O2",
  "te_fp8_format": "HYBRID",
  "te_amax_history_len": 32,
  "te_amax_compute_algo": "max",
  "use_generated_solution": true,
  "solution_field": "generated_solution",
  "use_wandb": true,
  "wandb_project": "qwen-math-fp8",
  "wandb_entity": null,
  "wandb_run_name": null,
  "wandb_tags": [
    "fp8",
    "qwen",
    "math"
  ],
  "wandb_notes": "",
  "wandb_resume": false,
  "wandb_watch_model": false,
  "wandb_watch_freq": 1000,
  "wandb_log_freq": 10,
  "wandb_log_model": false
}