paddlepaddle--paddlenlp
wehub-resource-sync
74 行
2.1 KiB
JSON
74 行
2.1 KiB
JSON
{
|
|
"logging_steps": 1,
|
|
"logging_dir": "vdl_log",
|
|
"train_datasets": "ppo-kk/34567ppl/train.jsonl",
|
|
"eval_datasets": "ppo-kk/5ppl/test.jsonl",
|
|
"prompt_key": "src",
|
|
"response_key": "tgt",
|
|
"actor_model_name_or_path": "meta-llama/Meta-Llama-3.1-8B",
|
|
"reward_model_name_or_path": "",
|
|
"use_rm_server": true,
|
|
"reward_server": "http://127.0.0.1:8731",
|
|
"rl_algorithm": "grpo",
|
|
"bf16": 1,
|
|
"fp16_opt_level": "O2",
|
|
"do_train": 1,
|
|
"adam_beta1": 0.9,
|
|
"adam_beta2": 0.95,
|
|
"dataloader_drop_last": 0,
|
|
"gradient_accumulation_steps": 32,
|
|
"ignore_save_lr_and_optim": 1,
|
|
"kl_loss_coeff": 0.001,
|
|
"learning_rate": 5e-7,
|
|
"lr_scheduler_type": "cosine",
|
|
"max_grad_norm": 1.0,
|
|
"max_length": 4608,
|
|
"max_steps": 3600,
|
|
"min_learning_rate": 2e-7,
|
|
"num_train_epochs": 1,
|
|
"output_dir": "checkpoints/llama-grpo",
|
|
"per_device_train_batch_size": 8,
|
|
"save_steps": 400,
|
|
"save_strategy": "steps",
|
|
"save_total_limit": 5,
|
|
"update_iters": 1,
|
|
"warmup_ratio": 0.03,
|
|
"weight_decay": 0.01,
|
|
"do_eval": 1,
|
|
"evaluation_strategy": "steps",
|
|
"per_device_eval_batch_size": 4,
|
|
"eval_steps": 10,
|
|
"eval_mode": "",
|
|
"max_dec_len": 4096,
|
|
"max_prompt_len": 512,
|
|
"min_dec_len": 1,
|
|
"normalize_advantage": 1,
|
|
"normalize_reward": 1,
|
|
"rollout_n": 8,
|
|
"per_device_prompt_batch_size": 32,
|
|
"repetition_penalty": 1.0,
|
|
"temperature": 0.7,
|
|
"top_p": 1.0,
|
|
"fused_linear": 1,
|
|
"recompute": 1,
|
|
"recompute_granularity": "full",
|
|
"recompute_use_reentrant": 1,
|
|
"use_flash_attention": 1,
|
|
"use_fused_head_and_loss_fn": 0,
|
|
"use_fusemt": 1,
|
|
"offload_level": "freeze_model optimizer train_model",
|
|
"release_grads": 1,
|
|
"clip_range_ratio": 0.2,
|
|
"clip_range_score": 10.0,
|
|
"clip_range_value": 5.0,
|
|
"ptx_coeff": 16.0,
|
|
"disable_tqdm": 1,
|
|
"sharding_parallel_degree": 1,
|
|
"sharding": "stage1",
|
|
"tensor_parallel_degree": 8,
|
|
"tensor_parallel_output": 0,
|
|
"pipeline_parallel_degree": 1,
|
|
"pipeline_parallel_config": "disable_p2p_cache_shape",
|
|
"sequence_parallel": 0,
|
|
"seed": 23
|
|
} |