| # 论文训练参数+正常推理参数 | |
| ### model | |
| model_name_or_path: Qwen/Qwen2.5-Coder-14B-Instruct | |
| trust_remote_code: true | |
| ### method | |
| stage: sft | |
| do_train: true | |
| finetuning_type: full | |
| deepspeed: /home/xuye_liu/yubo/LLaMA-Factory/examples/deepspeed/ds_z3_config.json | |
| ### dataset | |
| dataset: r2egym_sft_trajectories | |
| dataset_dir: /home/xuye_liu/yubo/LLaMA-Factory/data | |
| template: qwen | |
| cutoff_len: 20000 | |
| max_samples: 100000 | |
| overwrite_cache: true | |
| preprocessing_num_workers: 16 | |
| ### output | |
| output_dir: /home/xuye_liu/yubo/LLaMA-Factory/saves/R2EGym-14B-Agent-Coder-Instruct | |
| logging_steps: 10 | |
| resume_from_checkpoint: null | |
| save_steps: 200 | |
| plot_loss: true | |
| overwrite_output_dir: false | |
| ### train | |
| flash_attn: fa2 | |
| enable_liger_kernel: true | |
| use_unsloth_gc: true | |
| per_device_train_batch_size: 1 | |
| # Global batch size = per_device_train_batch_size * gradient_accumulation_steps * world_size. | |
| # Using GPUs 4,5,6,7 => world_size=4, so 1 * 2 * 4 = 8. | |
| gradient_accumulation_steps: 2 | |
| learning_rate: 1.0e-5 | |
| weight_decay: 0.05 | |
| num_train_epochs: 2.0 | |
| lr_scheduler_type: cosine | |
| warmup_ratio: 0.1 | |
| bf16: true | |
| ddp_timeout: 180000000 | |
| ### wandb | |
| report_to: none | |
| run_name: R2EGym-14B-Agent-Coder | |