| ### model | |
| model_name_or_path: /home/tianqiu/llama_3.1_sft | |
| trust_remote_code: true | |
| ## 加速 | |
| flash_attn: fa2 | |
| enable_liger_kernel: true | |
| # use_unsloth: true | |
| ### method | |
| stage: dpo | |
| do_train: true | |
| finetuning_type: lora | |
| lora_rank: 8 | |
| # lora_target: all | |
| lora_target: [q_proj, k_proj, v_proj, embed_tokens] | |
| pref_beta: 0.1 | |
| pref_loss: sigmoid # choices: [sigmoid (dpo), orpo, simpo] | |
| deepspeed: /home/tianqiu/LLaMA-Factory/examples/deepspeed/ds_z3_offload_config.json | |
| gradient_checkpointing: True | |
| ### dataset | |
| dataset: llama3.1_all_binary_0.5_dpo | |
| template: llama3 | |
| cutoff_len: 5000 | |
| max_samples: 58111 | |
| overwrite_cache: true | |
| preprocessing_num_workers: 16 | |
| dataloader_num_workers: 4 | |
| ### output | |
| output_dir: /home/tianqiu/llama_3.1_dpo_rank4 | |
| logging_steps: 10 | |
| save_steps: 500 | |
| plot_loss: true | |
| overwrite_output_dir: false | |
| save_only_model: false | |
| ### train | |
| per_device_train_batch_size: 1 | |
| gradient_accumulation_steps: 8 | |
| learning_rate: 5.0e-6 | |
| num_train_epochs: 3.0 | |
| lr_scheduler_type: cosine | |
| warmup_ratio: 0.1 | |
| bf16: true | |
| ddp_timeout: 180000000 | |
| resume_from_checkpoint: null | |
| # optim: paged_adamw_8bit 报错 | |
| ### eval | |
| # eval_dataset: dpo_en_demo | |
| # val_size: 0.1 | |
| # per_device_eval_batch_size: 1 | |
| # eval_strategy: steps | |
| # eval_steps: 500 | |