tvp4
收藏资源简介：
Mixture-of-Thoughts是一个由350k个经过验证的推理轨迹组成的精选数据集，这些轨迹是从DeepSeek-R1中提取的。该数据集涵盖了数学、编码和科学等领域的任务，旨在教会语言模型逐步推理。它被用于Open R1项目，以训练OpenR1-Distill-7B，这是一个复制了DeepSeek-R1-Distill-Qwen-7B推理能力的SFT模型。
Mixture-of-Thoughts is a curated dataset consisting of 350,000 verified reasoning traces extracted from DeepSeek-R1. This dataset covers tasks across domains including mathematics, coding, and science, aiming to teach language models to perform step-by-step reasoning. It has been used in the Open R1 project to train OpenR1-Distill-7B, an SFT model that replicates the reasoning capabilities of DeepSeek-R1-Distill-Qwen-7B.
数据集概述
基本信息
- 数据集名称: Mixture-of-Thoughts
- 任务类别: 文本生成
- 语言: 英语
- 数据集大小: >1TB
- 下载大小: 3.08GB (all配置)
- 数据集主页: https://huggingface.co/datasets/hiepp2/tvp4
数据集组成
-
配置:
all: 包含所有领域的数据code: 编程领域数据math: 数学领域数据science: 科学领域数据
-
数据量:
all: 349,317条推理轨迹code: 83,070条推理轨迹math: 93,733条推理轨迹science: 172,514条推理轨迹
数据特征
- 通用特征:
messages: 包含content(字符串)和role(字符串)的列表num_tokens: int64类型source: 字符串类型
数据来源
- math: 来自open-r1/OpenR1-Math-220k的
default子集 - code: 来自open-r1/codeforces-cots的
solutions和solutions_w_editorials子集 - science: 来自nvidia/Llama-Nemotron-Post-Training-Dataset的
science子集
使用方法
python from datasets import load_dataset
加载全部数据
dataset = load_dataset("open-r1/Mixture-of-Thoughts", "all", split="train")
加载特定领域数据
dataset_math = load_dataset("open-r1/Mixture-of-Thoughts", "math", split="train")
引用信息
bibtex @misc{openr1, title = {Open R1: A fully open reproduction of DeepSeek-R1}, url = {https://github.com/huggingface/open-r1}, author = {Hugging Face}, month = {January}, year = {2025} }
