| bf16: true | |
| cutoff_len: 131072 | |
| dataset: radiology_dataset | |
| dataset_dir: data | |
| ddp_timeout: 180000000 | |
| do_train: true | |
| eval_steps: 150 | |
| eval_strategy: steps | |
| finetuning_type: full | |
| flash_attn: auto | |
| gradient_accumulation_steps: 1 | |
| include_num_input_tokens_seen: true | |
| learning_rate: 5.0e-05 | |
| logging_steps: 5 | |
| lr_scheduler_type: cosine | |
| max_grad_norm: 1.0 | |
| max_samples: 100000 | |
| model_name_or_path: meta-llama/Llama-3.2-1B-Instruct | |
| num_train_epochs: 3.0 | |
| optim: adamw_torch | |
| output_dir: saves/Llama-3.2-1B-Instruct/full/llama3.2_1binst_full_sft | |
| packing: false | |
| per_device_eval_batch_size: 16 | |
| per_device_train_batch_size: 16 | |
| plot_loss: true | |
| preprocessing_num_workers: 16 | |
| report_to: all | |
| save_steps: 150 | |
| stage: sft | |
| template: llama3 | |
| val_size: 0.15 | |
| warmup_steps: 0 | |