This dataset is a version of tasksource/PRM800K prepared to fine tune a PRM model using TRL.
To replicate the dataset creation, go to main.py script in the files directory.
Following the guide to the data at openai/prm800k repository, it contains for each prompt, the list
of completions, where each one was obtained by exctracting the text from the list of steps, and for the cases with multiple alternative steps, a new completion was created.
Sample row:… See the full description on the dataset page:
https://huggingface.co/datasets/HuggingFaceH4/prm800k-trl-dedup.
