bitsydarel/KIMI-K2.5-1000000x

KIMI-K2.5-1000000x是一个包含1,000,000个推理轨迹的数据集，这些轨迹是从KIMI-K2.5模型蒸馏而来，专注于高推理任务。数据集规模为100K到1M之间，总token数达到50亿。它主要用于文本生成和问答任务，标签包括推理、思维链、指令调优和SFT。数据分布涵盖多个领域：编码占50%（包括Web开发、Python、C++、Java、JavaScript、C、Ruby、Lua、Rust和C#），科学占20%（物理、化学、生物），数学占15%（代数、微积分、概率），计算机科学占5%，逻辑问题占5%，创意写作占5%。此外，还包含一个多语言STEM子集，有100,000个完成项。数据集使用Apache 2.0许可证，语言为英语，通过修改的Datagen工具收集，耗时约80小时。

KIMI-K2.5-1000000x is a dataset containing 1,000,000 reasoning traces distilled from the KIMI-K2.5 model, focusing on high reasoning tasks. The dataset falls within the size category of 100K to 1M, with a total token count of 5 billion. It is primarily used for text-generation and question-answering tasks, with tags including reasoning, chain-of-thought, instruction-tuning, and SFT. The data distribution covers multiple domains: coding accounts for 50% (including web development, Python, C++, Java, JavaScript, C, Ruby, Lua, Rust, and C#), science for 20% (physics, chemistry, biology), math for 15% (algebra, calculus, probability), computer science for 5%, logical questions for 5%, and creative writing for 5%. Additionally, there is a MultilingualSTEM subset with 100,000 completions. The dataset is licensed under Apache 2.0, in English, and was collected using a modified Datagen tool over approximately 80 hours.