| bf16: true | |
| cutoff_len: 1024 | |
| dataset: cangjie | |
| dataset_dir: data | |
| ddp_timeout: 180000000 | |
| do_train: true | |
| eval_steps: 5000 | |
| eval_strategy: steps | |
| finetuning_type: lora | |
| flash_attn: auto | |
| gradient_accumulation_steps: 8 | |
| include_num_input_tokens_seen: true | |
| learning_rate: 0.0001 | |
| logging_steps: 2 | |
| lora_alpha: 32 | |
| lora_dropout: 0.1 | |
| lora_rank: 8 | |
| lora_target: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj | |
| loraplus_lr_ratio: 16 | |
| lr_scheduler_type: cosine | |
| max_grad_norm: 1.0 | |
| max_samples: 100000 | |
| model_name_or_path: Qwen/Qwen2-7B-Instruct | |
| num_train_epochs: 3.0 | |
| optim: adamw_torch | |
| output_dir: saves/Qwen2-7B-Chat/lora/train_08-16-2 | |
| packing: true | |
| per_device_eval_batch_size: 4 | |
| per_device_train_batch_size: 4 | |
| plot_loss: true | |
| preprocessing_num_workers: 16 | |
| report_to: none | |
| save_steps: 5000 | |
| stage: pt | |
| template: qwen | |
| val_size: 0.05 | |
| warmup_steps: 2 | |