five

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

收藏
Hugging Face2026-05-25 更新2026-05-31 收录
下载链接:
https://hf-mirror.com/datasets/iAmBoosted/gpt-oss-20b-reasoning-traces
下载链接
链接失效反馈
官方服务:
资源简介:
--- 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.
提供机构:
iAmBoosted
二维码
社区交流群
二维码
科研交流群
商业服务