Tachibana2-DeepSeek-R1-PREVIEW

**This is a preview of the full Tachibana 2 high-difficulty code-reasoning dataset**, containing the first ~6k rows. All responses generated by [deepseek-ai/DeepSeek-R1.](https://huggingface.co/deepseek-ai/DeepSeek-R1)

The full dataset will be released for everyone once it's ready!

This dataset contains:

- 6k high-difficulty synthetic code-reasoning prompts created by [Llama 3.1 405b Instruct](meta-llama/Llama-3.1-405B-Instruct), with an emphasis on task complexity and technical skill.
- Responses demonstrate the reasoning capabilities of DeepSeek's 685b parameter R1 reasoning model.

**Responses have not been filtered or edited at all:** the Tachibana 2 dataset strives to accurately represent the R1 model. Potential issues may include inaccurate answers and infinite thought loops. Tachibana 2 is presented as-is to be used at your discretion.

Users should consider applying their own sub-filtering and manual examination of the dataset before use in training.

Do as you will.

**本预览为完整Tachibana 2高难度代码推理数据集的预览版，包含前约6000条数据条目。所有回复均由[deepseek-ai/DeepSeek-R1](https://huggingface.co/deepseek-ai/DeepSeek-R1)生成。

完整数据集准备就绪后将向全体用户公开发布。

本数据集包含：
- 约6000条由[Llama 3.1 405B Instruct](meta-llama/Llama-3.1-405B-Instruct)生成的高难度合成代码推理提示词，该数据集侧重任务复杂度与技术熟练度；
- 其生成的回复可体现DeepSeek研发的6850亿参数R1推理模型的推理能力。

**所有回复均未经过任何过滤或编辑：Tachibana 2数据集旨在精准还原R1模型的原生输出表现。该数据集可能存在回复不准确、思维循环无限等问题。本数据集将按原始状态提供，使用者可自行决定使用方式。**

使用者在将该数据集用于模型训练前，应考虑自行进行次级筛选与人工核查。

请按需使用。