GeneralThought-195K

<img src="https://cdn-media.gr.inc/logoonly.png" alt="General Reasoning" width="150">

# GeneralThought-195K

**NEWEST RELEASE WITH 323K TRACES IS [HERE](https://huggingface.co/datasets/GeneralReasoning/GeneralThought-323K)**

> Thought wants to be free

Open reasoning data from the [General Reasoning](https://gr.inc) resource for March 3 2025.

The dataset contains questions, reference answers, reasoning traces, final answers and other metadata from several popular reasoning models including [DeepSeek-R1](https://gr.inc/DeepSeek/models/DeepSeek-R1/), [DeepSeek-R1-Zero](https://gr.inc/DeepSeek/models/DeepSeek-R1-Zero/), [OpenThoughts-32B](https://gr.inc/open-thoughts/models/OpenThinker-32B/), [LIMO](https://gr.inc/GAIR-NLP/models/LIMO/), [deepseek-r1-distill-llama-70b](https://gr.inc/DeepSeek/models/deepseek-r1-distill-llama-70b/), [DeepHermes-3-Llama-3-8B-Preview](https://gr.inc/NousResearch/models/DeepHermes-3-Llama-3-8B-Preview/) and [DeepScaleR-1.5B-Preview](https://gr.inc/agentica-org/models/DeepScaleR-1.5B-Preview/). We also include final answers from [o3-mini-2025-01-31](https://gr.inc/OpenAI/models/o3-mini-2025-01-31/), [gemini-2-flash-thinking-exp-01-21](https://gr.inc/Google/models/gemini-2-flash-thinking-exp-01-21/) and [claude-3-7-sonnet-20250219](https://gr.inc/Anthropic/models/claude-3-7-sonnet-20250219/) for comparison and evaluation. This release has 195,054 rows of data.

## Improvements

The main improvement in this update is trace diversity. There are many more reasoning traces beyond mathematics and code, including the natural sciences, humanities, social sciences, and general conversations.

## Metadata

A row of data contains the following information:

```python
row = {
'question_id': '296582', # question ID on the GR resource
'question_url': 'https://gr.inc/question/of-a-quadrilateral-if-its-angle-measures-are-in-the-ratio-of-3456-find-the-m', # URL on gr.inc
'question': 'Of a quadrilateral if its angle measures are in the ratio of 3:4:5:6, find the measure of each angle.', # Question text
'prev_messages': None, # previous messages in the conversation
'reference_answer': '60°, 80°, 100°, 120°', # Reference answer
'model_name': 'DeepSeek/DeepSeek-R1', # The model that generated the trace
'model_answer': 'The measures of the angles in the quadrilateral are calculated as follows...', # the model final answer text
'model_reasoning': 'Okay, so I need to find the measures of each angle in a quadrilateral...' # the model reasoning text
'task': 'Applying Ratios to Angle-Measure Sums', # name of the task on GR
'question_license': 'MIT', # license of the question
'question_source': 'General/VNet', # original dataset source or author on GR
'community_question_score': 0 # community score for the question on GR; negative means downvoted, positive upvoted,
'community_answer_score': 0, # community score for the answer on GR; negative means downvoted, positive upvoted
'verifier_score': 1.0 # an average verification score between 0-1; if multiple verifiers, this could be between, e.g. 0.5 if one verifier marks as correct, another incorrect
}
```

## How can I use the data?

The dataset is a great complement to [OpenThoughts-114k](https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k), [OpenR1](https://huggingface.co/datasets/open-r1/OpenR1-Math-Raw), [SYNTHETIC-1](https://huggingface.co/datasets/PrimeIntellect/SYNTHETIC-1), among others!

Here's some example use cases for our dataset:

- Perform SFT distillation and use it to train a small reasoning model.
- Ablate alongside datasets from our open source friends (and see if cross-group diversity helps).
- Analyse reasoning differences between models: reasoning length, language switching, and use of connectors like "wait" and "alternatively".

The verification side of the GR resource is still early, so we would recommend focusing on distillation (and rejection sampling) rather than online RL for now.

## Thanks

Thank you to the contributors of questions for this dataset:

First - thanks to the questions we sourced from folks like [Numina](https://huggingface.co/datasets/AI-MO/NuminaMath-CoT), [SCP-116k](https://huggingface.co/datasets/EricLu/SCP-116K), [natural_reasoning](https://huggingface.co/datasets/facebook/natural_reasoning) and others! We've credited you in the question_source field of each row of the dataset.

Thanks to GR community contributors who contributed:
- Jarius, otaldohenrikk, knight_raider, supahao, alpam, Esac, gonros, tomsercu, ryan, sidoneytemporary977, panpan, Tim_tom_0, arpitg1991, Doge, tginart, pcpthm, eli5, yych, caijie, yuchen.zhang2003, lockon, susheelt, wangxinjing, duyiyang, Slimane, FABGYUXIN, chendarcy, Sin, robintan, imhillxtz, navinahc, z, zhangdapao, yixiangRDS500

Going forward we will continue to credit those who contribute questions in future data dumps on Hugging Face 🤗.

**We will look to publish a paper with co-authorship for contributors.**

<img src="https://cdn-media.gr.inc/logoonly.png" alt="通用推理（General Reasoning）" width="150">

# GeneralThought-195K

**本次发布的含323K条推理轨迹的最新版本已上线[此处](https://huggingface.co/datasets/GeneralReasoning/GeneralThought-323K)**

> 思想本应自由

本数据集为2025年3月3日发布的[通用推理（General Reasoning）](https://gr.inc)资源下的开放推理数据。

本数据集涵盖多款主流推理模型生成的问题、参考答案、推理轨迹、最终答案及其他元数据，涉及模型包括[DeepSeek-R1](https://gr.inc/DeepSeek/models/DeepSeek-R1/)、[DeepSeek-R1-Zero](https://gr.inc/DeepSeek/models/DeepSeek-R1-Zero/)、[OpenThoughts-32B](https://gr.inc/open-thoughts/models/OpenThinker-32B/)、[LIMO](https://gr.inc/GAIR-NLP/models/LIMO/)、[deepseek-r1-distill-llama-70b](https://gr.inc/DeepSeek/models/deepseek-r1-distill-llama-70b/)、[DeepHermes-3-Llama-3-8B-Preview](https://gr.inc/NousResearch/models/DeepHermes-3-Llama-3-8B-Preview/)及[DeepScaleR-1.5B-Preview](https://gr.inc/agentica-org/models/DeepScaleR-1.5B-Preview/)。此外，为便于对比与评估，本数据集还纳入了[o3-mini-2025-01-31](https://gr.inc/OpenAI/models/o3-mini-2025-01-31/)、[gemini-2-flash-thinking-exp-01-21](https://gr.inc/Google/models/gemini-2-flash-thinking-exp-01-21/)与[claude-3-7-sonnet-20250219](https://gr.inc/Anthropic/models/claude-3-7-sonnet-20250219/)的最终答案。本次发布共包含195,054条数据行。

## 改进亮点

本次更新的核心改进在于推理轨迹的多样性。本数据集新增了大量覆盖数学与代码之外领域的推理轨迹，涵盖自然科学、人文社科及日常对话等场景。

## 元数据格式

单条数据行包含以下字段：

python
row = {
'question_id': '296582', # GR资源上的问题ID
'question_url': 'https://gr.inc/question/of-a-quadrilateral-if-its-angle-measures-are-in-the-ratio-of-3456-find-the-m', # gr.inc平台上的问题链接
'question': '已知四边形的四个内角之比为3:4:5:6，求各内角的度数。', # 问题文本
'prev_messages': None, # 对话历史消息
'reference_answer': '60°, 80°, 100°, 120°', # 参考答案
'model_name': 'DeepSeek/DeepSeek-R1', # 生成该推理轨迹的模型
'model_answer': '该四边形各内角的计算过程如下...', # 模型最终答案文本
'model_reasoning': '好的，我需要先求出四边形各内角的度数...' # 模型推理过程文本
'task': '角度比例与内角和应用', # GR资源上的任务名称
'question_license': 'MIT', # 问题授权协议
'question_source': 'General/VNet', # GR资源上的原始数据集来源或作者
'community_question_score': 0 # GR资源上的社区问题评分；负值代表差评，正值代表好评
'community_answer_score': 0, # GR资源上的社区答案评分；负值代表差评，正值代表好评
'verifier_score': 1.0 # 0-1区间的平均验证得分；若存在多位验证者，得分可能介于区间内，例如0.5代表一位验证者判定正确、另一位判定错误
}

## 数据集使用场景

本数据集可作为[OpenThoughts-114k](https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k)、[OpenR1](https://huggingface.co/datasets/open-r1/OpenR1-Math-Raw)、[SYNTHETIC-1](https://huggingface.co/datasets/PrimeIntellect/SYNTHETIC-1)等数据集的优质补充！

本数据集的典型应用场景包括：
- 开展监督微调（Supervised Fine-Tuning, SFT）蒸馏，以训练小型推理模型
- 与开源伙伴的数据集开展消融实验，探究跨群体多样性对模型性能的影响
- 分析不同模型间的推理差异：如推理长度、语言切换习惯，以及“等一下”“或者”等连接词的使用频率

目前GR资源的验证模块仍处于早期阶段，因此我们建议现阶段优先关注蒸馏（及拒绝采样）技术，而非在线强化学习（RL）。

## 致谢

感谢为本数据集贡献问题的所有参与者：

首先，感谢我们从[Numina](https://huggingface.co/datasets/AI-MO/NuminaMath-CoT)、[SCP-116k](https://huggingface.co/datasets/EricLu/SCP-116K)、[natural_reasoning](https://huggingface.co/datasets/facebook/natural_reasoning)等来源获取的问题贡献者！我们已在每条数据行的`question_source`字段中为您署名。

同时感谢GR社区的以下贡献者：
Jarius, otaldohenrikk, knight_raider, supahao, alpam, Esac, gonros, tomsercu, ryan, sidoneytemporary977, panpan, Tim_tom_0, arpitg1991, Doge, tginart, pcpthm, eli5, yych, caijie, yuchen.zhang2003, lockon, susheelt, wangxinjing, duyiyang, Slimane, FABGYUXIN, chendarcy, Sin, robintan, imhillxtz, navinahc, z, zhangdapao, yixiangRDS500

未来我们将继续在Hugging Face 🤗的后续数据发布中为所有问题贡献者署名。

**我们计划为所有贡献者提供论文共同作者署名资格。**