| ### model | |
| model_name_or_path: meta-llama/Meta-Llama-3.1-8B-Instruct | |
| new_special_tokens: <design_start>,<design_end>,<design_body>,<molecule>,<retro_start>,<retro_end>,<retro_body>,<rollback_start>,<rollback_end> | |
| graph_decoder_path: saves/graph_decoder | |
| graph_encoder_path: saves/graph_encoder | |
| graph_predictor_path: saves/graph_predictor | |
| graph_lm_connector_path: null #specify when resume | |
| ### method | |
| stage: mmsft | |
| do_train: true | |
| finetuning_type: lora | |
| lora_target: all | |
| flash_attn: disabled | |
| learned_query_size: 8 | |
| ### dataset | |
| dataset: molqa_train_examples | |
| template: llama3 | |
| cutoff_len: 2048 | |
| overwrite_cache: true | |
| preprocessing_num_workers: 16 | |
| ### output | |
| output_dir: saves/Llama-3.1-8B-Instruct-Adapter | |
| logging_steps: 10 | |
| save_steps: 2000 | |
| plot_loss: true | |
| overwrite_output_dir: true | |
| ### to resume | |
| # overwrite_output_dir: false | |
| ### train | |
| per_device_train_batch_size: 10 | |
| gradient_accumulation_steps: 2 | |
| learning_rate: 1.0e-4 | |
| num_train_epochs: 4.0 | |
| lr_scheduler_type: cosine | |
| warmup_ratio: 0.1 | |
| ddp_timeout: 180000000 | |
| bf16: true | |
| pure_bf16: true | |
| ### train loss | |
| loss_weight_retro: 1 | |
| loss_weight_design: 1 | |
| loss_weight_lm: 1 | |
| ### eval | |
| val_size: 0.1 | |
| per_device_eval_batch_size: 6 | |
| eval_strategy: steps | |
| eval_steps: 2000 | |
| ### specify if connected to wandb | |
| report_to: 'none' |