---
pretty_name: OpenReasoning Mixed 100K
language:
- en
task_categories:
- text-generation
tags:
- reasoning
- math
- code
- science
configs:
- config_name: default
  data_files:
  - split: train
    path: train.parquet
---

# OpenReasoning Mixed 100K

This dataset is a 100,000-row prompt-only mixture prepared for reproducing the
Qwen3-1.7B on-policy distillation experiments described in
[arXiv:2607.15161](https://arxiv.org/abs/2607.15161).

This is an independent reproduction artifact, not an official dataset release
from the paper authors.

## Composition

| Domain | Rows | Upstream dataset | Config / split |
|---|---:|---|---|
| Math | 33,334 | `nvidia/OpenMathReasoning` | `default / cot` |
| Science | 33,333 | `nvidia/OpenScienceReasoning-2` | `default / train` |
| Code | 33,333 | `nvidia/OpenCodeReasoning` | `split_0 / split_0` |

Only question text and provenance metadata are retained. Upstream answers and
reasoning traces are not included.

## Fields

- `messages`: one-message chat record containing the user question
- `domain`: `math`, `science`, or `code`
- `source_dataset`: upstream Hugging Face dataset ID
- `source_config`: upstream dataset config
- `source_split`: upstream split
- `source_shard`: sampled upstream Parquet shard index
- `source_id`: stable source-row/problem identifier used by the builder
- `prompt_sha256`: SHA-256 of the prompt text

## Construction

- Seed: `42`
- Sampling: balanced allocation over randomly ordered Parquet shards, followed
  by random row-group and row selection within each shard
- Maximum selected source shards per domain: `12`
- Quality filtering: none
- Length filtering: none
- Prompt deduplication: none
- Answers and reasoning traces retained: no

The exact construction manifest is included as `manifest.json`.

## Duplicate prompts

The dataset contains 74,644 unique prompt hashes and 25,356 duplicate prompt
rows. Most duplicates come from upstream reasoning datasets containing multiple
solution traces for the same underlying question. They are intentionally
preserved to match the row-sampling interpretation used by this reproduction.

In particular, the code portion contains 33,333 rows but 10,538 unique prompts.
Users who require unique problems should deduplicate using `prompt_sha256`.

## Upstream revisions

- `nvidia/OpenMathReasoning`:
  `d3d08664755704f422af97d43a7ff0ded4bd95df`
- `nvidia/OpenScienceReasoning-2`:
  `174b02c9cdf231f220765b2a1d5ece4550921894`
- `nvidia/OpenCodeReasoning`:
  `20a1ca19c0d050fe9057fc08339d6b370ec1c67a`

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("YangyiH/openreasoning_mixed_100k", split="train")
```

## Licensing and attribution

This repository redistributes prompt text derived from the three upstream
NVIDIA datasets listed above. Review and comply with each upstream dataset card,
license, terms, and source attribution requirements before use or
redistribution. No single new license is asserted here over upstream content.

## Limitations

- The paper authors have not released the exact data-mixing implementation.
- The mixture preserves duplicate prompts and should not be interpreted as
  100,000 unique questions.
- No additional quality, difficulty, contamination, or prompt-length filtering
  was applied.
- The mixture has not been audited for all possible benchmark overlap or
  sensitive content inherited from upstream sources.