| backdoor_evals: | |
| - jailbreak | |
| batch_size: 8 | |
| compute_ppl: false | |
| evaluate_model_performance: false | |
| evaluate_model_performance_at_the_end: false | |
| folder_name: null | |
| ft_datasets: | |
| - !!python/object/apply:src.data.dataset.DatasetType | |
| - OpenMathInstruct | |
| - !!python/object/apply:src.data.dataset.DatasetType | |
| - AlpacaGPT4 | |
| - !!python/object/apply:src.data.dataset.DatasetType | |
| - CodeAlpaca | |
| - !!python/object/apply:src.data.dataset.DatasetType | |
| - PubMedQA | |
| lora_config: null | |
| max_new_tokens: 100 | |
| metadatas: null | |
| min_new_tokens: 10 | |
| n_samples: 1500 | |
| oversample: 1 | |
| ppl_model: meta-llama/Llama-3.1-8B-Instruct | |
| prompt_datasets: | |
| - data_fields: | |
| - instruction | |
| - output | |
| path: Grogros/jailbreak_test | |
| split: train | |
| prompt_length: 50 | |
| save_model: true | |
| sequence_length: 512 | |
| skip_if_exists: false | |
| streaming: true | |
| temperature: 1.0 | |
| training_args: | |
| bf16: false | |
| do_train: true | |
| fp16: false | |
| gradient_accumulation_steps: 16 | |
| gradient_checkpointing: false | |
| learning_rate: 5.0e-05 | |
| max_steps: 2000 | |
| num_train_epochs: 1 | |
| optim: adafactor | |
| output_dir: Grogros/Qwen2-7B-Instruct-TRAINOpenMathInstruct | |
| overwrite_output_dir: true | |
| per_device_train_batch_size: 2 | |
| push_to_hub: true | |
| report_to: tensorboard | |
| save_steps: 500 | |
| use_tmp: false | |