| bf16: true | |
| create_new_adapter: true | |
| cutoff_len: 512 | |
| dataset: top_5_training_dataset | |
| dataset_dir: data | |
| ddp_timeout: 180000000 | |
| do_train: true | |
| finetuning_type: lora | |
| flash_attn: auto | |
| gradient_accumulation_steps: 5 | |
| include_num_input_tokens_seen: true | |
| label_smoothing_factor: 0.05 | |
| learning_rate: 2.0e-05 | |
| logging_steps: 25 | |
| lora_alpha: 64 | |
| lora_dropout: 0.05 | |
| lora_rank: 32 | |
| lora_target: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj | |
| loraplus_lr_ratio: 8 | |
| lr_scheduler_type: cosine_with_restarts | |
| max_grad_norm: 0.3 | |
| max_samples: 100000 | |
| model_name_or_path: google/gemma-3-12b-it | |
| num_train_epochs: 4.0 | |
| optim: adamw_torch | |
| output_dir: saves/Gemma-3-12B-Instruct/lora/train_2025-05-14-10-29-44 | |
| packing: false | |
| per_device_train_batch_size: 5 | |
| plot_loss: true | |
| preprocessing_num_workers: 16 | |
| report_to: none | |
| save_steps: 250 | |
| stage: sft | |
| template: alpaca | |
| trust_remote_code: true | |
| warmup_steps: 150 | |
| weight_decay: 0.01 | |