| seed: 101112 | |
| ### model | |
| model_name_or_path: meta-llama/Meta-Llama-3-8B-Instruct | |
| trust_remote_code: true | |
| flash_attn: auto | |
| use_cache: false | |
| ### method | |
| stage: sft | |
| do_train: true | |
| finetuning_type: p-tuning | |
| task_type: CAUSAL_LM | |
| num_virtual_tokens: 100 | |
| encoder_reparameterization_type: MLP | |
| encoder_hidden_size: 768 | |
| encoder_dropout: 0.1 | |
| ### dataset | |
| dataset: hellaswag | |
| template: llama3 | |
| cutoff_len: 2048 | |
| overwrite_cache: true | |
| preprocessing_num_workers: 8 | |
| dataloader_num_workers: 2 | |
| packing: false | |
| ### output | |
| output_dir: saves_multiple/p-tuning/llama-3-8b-instruct/train_hellaswag_101112_1760638082 | |
| logging_steps: 5 | |
| save_steps: 0.05 | |
| overwrite_output_dir: true | |
| save_only_model: false | |
| plot_loss: true | |
| include_num_input_tokens_seen: true | |
| push_to_hub: true | |
| push_to_hub_organization: rbelanec | |
| load_best_model_at_end: true | |
| save_total_limit: 1 | |
| ### train | |
| per_device_train_batch_size: 4 | |
| learning_rate: 1.0e-3 | |
| num_train_epochs: 20 | |
| weight_decay: 1.0e-5 | |
| lr_scheduler_type: cosine | |
| bf16: true | |
| ddp_timeout: 180000000 | |
| resume_from_checkpoint: null | |
| warmup_ratio: 0.1 | |
| optim: adamw_torch | |
| report_to: | |
| - wandb | |
| run_name: p-tuning_llama-3-8b-instruct_train_hellaswag_101112_1760638082 | |
| ### eval | |
| per_device_eval_batch_size: 4 | |
| eval_strategy: steps | |
| eval_steps: 0.05 | |
| val_size: 0.1 | |