To construct the initial supervised training data, we employ the MATH dataset as a foundation. Each solution is systematically reformatted into a structured step-by-step explanation using the gpt-4o-2024-08-06 model. The reformatting process ensures that each logical step is clearly delineated and separated… See the full description on the dataset page:
https://huggingface.co/datasets/Jing-Xun/HS-STaR.
