iAmBoosted/gpt-oss-20b-reasoning-traces

---
license: other
license_name: mixed-see-licensing-section
task_categories:
- text-generation
language:
- en
tags:
- reasoning
- chain-of-thought
- synthetic
- distillation
size_categories:
- 1K<n<10K
---

# GPT-OSS-20B Reasoning Traces

3,333 reasoning traces generated by [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) and filtered for clean, terminating reasoning. It was built to distill GPT-OSS's tight reasoning style into smaller models, and is the training set behind [iAmBoosted/Qwen3.5-9B-OSS-Distilled](https://huggingface.co/iAmBoosted/Qwen3.5-9B-OSS-Distilled).

## What's in it

Each record pairs a **prompt** with GPT-OSS-20B's **full reasoning trace and final answer**, in chat-message form, ready for supervised fine-tuning (SFT).

- ~4,000 raw traces were generated, then filtered down to **3,333** — traces that failed to terminate cleanly or were malformed were dropped.
- Language: English.

### Fields

| Field | Description |
|---|---|
| `id` | Unique record ID, prefixed with its source dataset |
| `domain` | Subject area: `math`, `code`, `biology`, `chemistry`, `physics`, or `puzzle` |
| `source` | The dataset the prompt was harvested from (see Licensing) |
| `messages` | Chat-format list: the user prompt, and the assistant's `<think>` reasoning trace + answer |
| `answer_status` | Whether the final answer could be verified: `correct`, `no_reference`, or `unverifiable` |
| `reasoning_words` | Word count of the reasoning trace |

### Breakdown by domain

| Domain | Records |
|---|---|
| math | 912 |
| code | 862 |
| biology | 457 |
| chemistry | 440 |
| physics | 434 |
| puzzle | 228 |

## How it was built

1. Prompts were harvested from four open datasets (see below).
2. GPT-OSS-20B generated a reasoning trace and answer for each prompt.
3. A filtering step removed traces that did not terminate cleanly, leaving 3,333 SFT-ready examples.

The filter checked for clean termination, **not** correctness. The `answer_status` field records what could be verified afterward: of the 3,333 traces, 556 were confirmed correct against a reference answer, 2,020 had no reference answer to check against, and 757 were unverifiable. Treat the traces as demonstrations of reasoning *style*, not as a verified answer key.

## Source datasets & licensing

The **traces** are outputs of GPT-OSS-20B (Apache-2.0). The **prompts** are redistributed from four datasets, so this dataset is **multi-licensed** — it is *not* uniformly Apache-2.0. Every record carries a `source` field, so you can filter by origin, and therefore by license.

| Source | Records | Domain | License |
|---|---|---|---|
| [open-thoughts/OpenThoughts-114k](https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k) | 2,730 | math, code, science | Apache-2.0 |
| [hivaze/LOGIC-701](https://huggingface.co/datasets/hivaze/LOGIC-701) | 228 | puzzle | Apache-2.0 |
| [AI-MO/NuminaMath-1.5](https://huggingface.co/datasets/AI-MO/NuminaMath-1.5) | 202 | math | Apache-2.0 |
| [nvidia/OpenCodeReasoning](https://huggingface.co/datasets/nvidia/OpenCodeReasoning) | 173 | code | predominantly CC-BY-4.0 (some Apache-2.0 / MIT, per sample) |

The 173 records with `source = nvidia/OpenCodeReasoning` are **predominantly CC-BY-4.0**, with a minority under Apache-2.0 or MIT. **CC-BY-4.0 requires attribution** — credit is given here to NVIDIA and the OpenCodeReasoning project, whose questions are themselves drawn from competitive-programming sources (TACO, APPS, CodeContests, Codeforces, and others). If you reuse those records, retain this attribution.

**Summary:** roughly 95% of the prompts (the OpenThoughts-114k, LOGIC-701, and NuminaMath-1.5 records) are Apache-2.0; the 173 OpenCodeReasoning records are CC-BY-4.0 (with a minority Apache-2.0 / MIT). Treat the dataset as multi-licensed and attribute accordingly.

## Limitations

- **Synthetic.** Traces are model-generated. They were filtered for clean *termination*; most were not verified for *correctness* (see `answer_status` above).
- English only.
- Reasoning-domain specific (math / code / science / logic) — not general-purpose text.

## Citation

If you use this dataset, please credit the four source datasets above (OpenThoughts-114k, LOGIC-701, NuminaMath-1.5, OpenCodeReasoning) and OpenAI's GPT-OSS-20B as the trace generator.