| --- |
| {} |
| --- |
| # allenai/open_instruct: Generation Dataset |
| |
| See https://github.com/allenai/open-instruct/blob/main/docs/algorithms/rejection_sampling.md for more detail |
|
|
| ## Configs |
|
|
| ``` |
| args: |
| {'add_timestamp': True, |
| 'dataset_end_idx': 2, |
| 'dataset_mixer_list': ['VGraf/no_safety_tulu_pref', '2'], |
| 'dataset_shuffle': True, |
| 'dataset_splits': ['train', 'train'], |
| 'dataset_start_idx': 0, |
| 'hf_entity': 'VGraf', |
| 'hf_repo_id': 'generation_multi', |
| 'include_persona': False, |
| 'max_qs': 7, |
| 'mode': 'generation', |
| 'model2_name_or_path': 'cleanrl/EleutherAI_pythia-6.9b-deduped__sft__tldr', |
| 'model_name_or_path': 'gpt-3.5-turbo-0125', |
| 'mt_token_cutoff': 8000, |
| 'num_turns': 1, |
| 'push_to_hub': True, |
| 'revision': 'main', |
| 'revision2': 'main', |
| 'save_filename': '/weka/oe-adapt-default/victoriag/synth_data/nonsafety_pref_related_prompts.jsonl', |
| 'skill': 'related_prompts', |
| 'user_type': 'follow-up'} |
| |
| dataset_config: |
| {'batched': False, |
| 'binary_messages_key': 'messages', |
| 'chat_template': 'simple_chat', |
| 'convert_preference_to_binary_dataset': False, |
| 'label': 'binary_labels', |
| 'load_from_cache_file': False, |
| 'max_prompt_token_lenth': None, |
| 'max_token_length': None, |
| 'ncols': 2, |
| 'num_proc': 1, |
| 'preference_chosen_key': 'chosen', |
| 'preference_rejected_key': 'rejected', |
| 'sanity_check': True, |
| 'sanity_check_max_samples': 100, |
| 'sft_messages_key': 'messages', |
| 'train_only_on_prompt': False} |
| |
| gen_args: |
| {'num_completions': 1, |
| 'response_length': 2048, |
| 'temperature': 0.8, |
| 'tensor_parallel_size': 1, |
| 'top_p': 0.9} |
| ``` |
|
|
| ## Reproduce this dataset |
|
|
| 1. Download the `['generation_multi.py', 'nonsafety_pref_related_prompts.jsonl']` from the https://huggingface.co/datasets/VGraf/generation_multi_1742346405. |
| 2. Run `python open_instruct/rejection_sampling/generation_multi.py --dataset_mixer_list VGraf/no_safety_tulu_pref 2 --dataset_splits train train --model_name_or_path gpt-3.5-turbo-0125 --num_completions 1 --dataset_shuffle --max_qs 7 --skill related_prompts --save_filename /weka/oe-adapt-default/victoriag/synth_data/nonsafety_pref_related_prompts.jsonl --sanity_check --push_to_hub` |
|
|
