---
license: apache-2.0
task_categories:
- text-generation
language:
- en
pretty_name: OpenR1-GNT-69k
size_categories:
- 10K<n<100K
---

# OpenR1-GNT-69k

This dataset contains 69,322 math reasoning training examples derived from OpenR1-Math-220k by keeping DeepSeek-R1 generations marked incorrect by math verification.

## Files

- `train.parquet`: main training split used by the GNT/LUFFY training scripts.

## Schema

- `data_source`: source tag from the original OpenR1 example.
- `prompt_neg`: system/user messages where the user prompt includes the problem and an incorrect teacher trace as a reference solution to avoid.
- `prompt_plain`: system/user messages with the original problem only.
- `reward_model`: rule-based reward metadata with the ground-truth answer.
- `ability`: ability tag, kept for compatibility with the training pipeline.
- `target`: incorrect teacher trace retained from OpenR1 generations.
- `extra_info`: auxiliary metadata.

## Source

This dataset is based on `open-r1/OpenR1-Math-220k`.