35 lines
897 B
JSON
35 lines
897 B
JSON
|
{
|
||
|
|
"model": "./math_grpo_hard",
|
||
|
|
"output_dir": "./math_grpo_hard2",
|
||
|
|
"max_problems": 2000,
|
||
|
|
"num_train_epochs": 1,
|
||
|
|
"per_device_train_batch_size": 1,
|
||
|
|
"gradient_accumulation_steps": 16,
|
||
|
|
"num_generations": 4,
|
||
|
|
"learning_rate": 5e-06,
|
||
|
|
"max_completion_length": 8192,
|
||
|
|
"mask_truncated_completions": true,
|
||
|
|
"min_difficulty": 7.0,
|
||
|
|
"max_difficulty": 10.0,
|
||
|
|
"vllm_max_model_length": null,
|
||
|
|
"seed": 42,
|
||
|
|
"resume_checkpoint": null,
|
||
|
|
"save_interval_hours": 1.0,
|
||
|
|
"reward_type": "accuracy",
|
||
|
|
"cos_correct_short": 1.0,
|
||
|
|
"cos_correct_long": 0.5,
|
||
|
|
"cos_wrong_short": -1.0,
|
||
|
|
"cos_wrong_long": 0.0,
|
||
|
|
"beta": 0.04,
|
||
|
|
"loss_type": "dapo",
|
||
|
|
"epsilon": 0.2,
|
||
|
|
"epsilon_high": 0.28,
|
||
|
|
"zero_variance_filter": false,
|
||
|
|
"use_vllm": true,
|
||
|
|
"vllm_gpu_memory_utilization": 0.3,
|
||
|
|
"use_lora": true,
|
||
|
|
"lora_r": 32,
|
||
|
|
"lora_alpha": 64,
|
||
|
|
"lora_dropout": 0.05,
|
||
|
|
"_timestamp": "2026-06-08 08:46:08"
|
||
|
|
}
|