o Ti@sddlmZmZddlmZddlmZddlmZddl m Z eGdddZ dd Z d d Z ed Zed krne e ZedZdddZededZeje ddgdejdZejrpeejejejdddSdSdS)) dataclassfield)Optional) load_dataset) ModelCard)HfArgumentParserc@s^eZdZUdZedddidZeed<edddidZe ed <ed dd idZ e e ed <d S) ScriptArgumentsa Arguments for the script. Args: push_to_hub (`bool`, *optional*, defaults to `False`): Whether to push the dataset to the Hugging Face Hub. repo_id (`str`, *optional*, defaults to `"trl-lib/prm800k"`): Hugging Face repository ID to push the dataset to. dataset_num_proc (`int` or `None`, *optional*, defaults to `None`): Number of workers to use for dataset processing. Fhelpz4Whether to push the dataset to the Hugging Face Hub.)defaultmetadata push_to_hubztrl-lib/prm800kz2Hugging Face repository ID to push the dataset to.repo_idNz0Number of workers to use for dataset processing.dataset_num_proc) __name__ __module__ __qualname____doc__rr bool__annotations__r strrrintrrM/home/ubuntu/.local/lib/python3.10/site-packages/examples/datasets/prm800k.pyrs  rc Cs8g}|dd}g}g}|ddD]~}|ddur(|ddur(|ddur(nit|dD]/\}}||dkr]|d}|dd|g} |d d k} |dd| g} ||| | d q.|ddurs|d|d} | d d k} n|ddur|d} d } nn| d}|||| q||||d |S) Nquestionproblemlabelsteps completionshuman_completionchosen_completiontextrating)promptrlabelsT) enumerateappend) exampleoutputsr#previous_completionsprevious_labelsstepcompletion_idx completioncontentrrr$rrrrprocess_example3s6 $      r/cs\gt|d}t|D]fdd|D}t|q fdddDS)Nrcsi|] \}}||qSrr).0kv)idxrr ^sz!process_batch..cs i|] fddDqS)csg|]}|qSrr)r0r2r1rr asz,process_batch...r)r0)r(r5rr4as r)lenrangeitemsextendr/)examples batch_sizer'r)r3r(r process_batchZs  r=a --- tags: [trl] --- # PRM800K Dataset ## Summary The PRM800K dataset is a processed version of [OpenAI's PRM800K](https://github.com/openai/prm800k), designed to train models using the [TRL library](https://github.com/huggingface/trl) for stepwise supervision tasks. It contains 800,000 step-level correctness labels for model-generated solutions to problems from the MATH dataset. This dataset enables models to learn and verify each step of a solution, enhancing their reasoning capabilities. ## Data Structure - **Format**: [Standard](https://huggingface.co/docs/trl/main/dataset_formats#standard) - **Type**: [Stepwise supervision](https://huggingface.co/docs/trl/main/dataset_formats#stepwise-supervision) Columns: - `"prompt"`: The problem statement. - `"completions"`: A list of reasoning steps generated to solve the problem. - `"labels"`: A list of booleans or floats indicating the correctness of each corresponding reasoning step. This structure allows models to learn the correctness of each step in a solution, facilitating improved reasoning and problem-solving abilities. ## Generation script The script used to generate this dataset can be found [here](https://github.com/huggingface/trl/blob/main/examples/datasets/prm800k.py). __main__zUhttps://github.com/openai/prm800k/raw/refs/heads/main/prm800k/data/phase1_train.jsonlzThttps://github.com/openai/prm800k/raw/refs/heads/main/prm800k/data/phase1_test.jsonl)traintestjson) data_filesT )labeler timestamp generationis_quality_control_questionis_initial_screening_questionrr)batchedr<remove_columnsnum_procdataset) repo_typeN) dataclassesrrtypingrdatasetsrhuggingface_hubr transformersrrr/r= model_cardrparserparse_args_into_dataclasses script_argsrBrLmaprr r rrrrs:    '