reasoning-v1-20m


We are excited to release a synthetic reasoning dataset containing 22mil+ general reasoning questions and responses generated using [deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B). While there have been multiple efforts to build open reasoning datasets for math and code tasks, we noticed a lack of large datasets containing reasoning traces for diverse non code/math topics like social and natural sciences, education, creative writing and general conversations, which is why we decided to release this dataset.<br>
*Note: Please note that in this instance we have not verified the reasoning traces and answers for accuracy.*

**Dataset details:**<br>
*Total number of rows:* 22.2 million rows<br>
*Total number of tokens:* 35.8 billion tokens

The dataset can be used to fine-tune smaller, more efficient models to mimic the reasoning capabilities of larger models like DeepSeek-R1 using SFT.

**Response format:**
```
<think>
-- reasoning trace --
</think>
-- answer --
```

**Loading the dataset:**
```python
from datasets import load_dataset
ds = load_dataset("glaiveai/reasoning-v1-20m", split="train")
```

我们很高兴发布一款合成推理数据集，其中包含超过2200万条通用推理问题及对应答案，由[deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B)生成。此前已有多项工作针对数学与代码任务构建开源推理数据集，但我们注意到，目前仍缺乏覆盖社会科学、自然科学、教育学、创意写作及通用对话等多元非代码/非数学主题的带推理轨迹的大型数据集，这也是我们发布本数据集的初衷。
*注：请注意，本次发布的数据集未对推理轨迹与答案的准确性进行验证。*

**数据集详情：**
*总行数：* 2220万行
*总Token（Token）数：* 358亿Token

本数据集可用于通过监督微调（Supervised Fine-Tuning，SFT）对更小、更高效的模型进行微调，使其具备类似DeepSeek-R1这类大模型的推理能力。

**回复格式：**

<think>
-- 推理轨迹 --
</think>
-- 答案 --

**加载数据集：**
python
from datasets import load_dataset
ds = load_dataset("glaiveai/reasoning-v1-20m", split="train")

搜集汇总

背景与挑战

背景概述

该数据集包含2200万条通用推理问题和回答，覆盖社会科学、自然科学等多个非数学/代码领域，总token数达35.8亿，用于训练小型模型模仿大型模型的推理能力。