| bf16: true | |
| cutoff_len: 2048 | |
| dataset: adpr_train | |
| dataset_dir: data | |
| ddp_timeout: 180000000 | |
| do_train: true | |
| eval_steps: 100 | |
| eval_strategy: steps | |
| finetuning_type: lora | |
| flash_attn: auto | |
| gradient_accumulation_steps: 8 | |
| include_num_input_tokens_seen: true | |
| learning_rate: 5.0e-05 | |
| logging_steps: 5 | |
| lora_alpha: 128 | |
| lora_dropout: 0.01 | |
| lora_rank: 64 | |
| lora_target: q_proj,v_proj,k_proj,o_proj,gate_proj,down_proj,up_proj | |
| lr_scheduler_type: cosine | |
| max_grad_norm: 1.0 | |
| max_samples: 100000 | |
| model_name_or_path: GreatCaptainNemo/ProLLaMA | |
| num_train_epochs: 3.0 | |
| optim: adamw_torch | |
| output_dir: saves/Custom/lora/train_2025-04-05-23-57-03 | |
| packing: false | |
| per_device_eval_batch_size: 16 | |
| per_device_train_batch_size: 16 | |
| plot_loss: true | |
| preprocessing_num_workers: 16 | |
| report_to: none | |
| resize_vocab: true | |
| save_steps: 100 | |
| stage: sft | |
| template: alpaca | |
| trust_remote_code: true | |
| val_size: 0.1 | |
| warmup_steps: 20 | |