Datasets:

rodriguescarson
/

adaption-math-step-verification

license: mit
pretty_name: Math Solution Step Verification
language:
  - en
task_categories:
  - text-classification
  - text-generation
tags:
  - adaption
  - autoscientist
  - sft
  - instruction-tuning
  - fine-tuning
  - math
  - process-reward
  - verification
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data.parquet
source_datasets:
  - openai/prm800k
  - HuggingFaceH4/prm800k-trl-dedup

Math Solution Step Verification

Judge whether a step in a worked maths solution is correct, from human step-level labels.


Rows	1,000
Domain	mathematics
Format	`data.parquet`, one row per example
Licence	mit
Built for	supervised fine-tuning (SFT) experiments on Adaption AutoScientist

Columns

Column	Description
`prompt`	The prompt (user turn) as uploaded.
`completion`	The target response as uploaded.
`enhanced_prompt`	Empty in this dataset.
`enhanced_completion`	Response after Adaption processing (rewrite or augmentation).
`reasoning_trace`	Step-by-step reasoning trace attached by Adaption's Adaptive Data.

How it was built

Built from PRM800K (via the TRL-deduplicated copy); the explained variant adds a written justification.

Sources and licence

PRM800K (MIT)
HuggingFaceH4/prm800k-trl-dedup (MIT)

Notes

Columns enhanced_prompt are empty in this dataset (the Adaption export reserves them for rewritten text).

Loading

from datasets import load_dataset
ds = load_dataset("rodriguescarson/adaption-math-step-verification", split="train")

import pandas as pd
df = pd.read_parquet("hf://datasets/rodriguescarson/adaption-math-step-verification/data.parquet")

Published by Carson Rodrigues (Hugging Face, Kaggle).