One Million Reasoning Traces: KIMI‑K2.5‑1000000x Dataset

The **KIMI‑K2.5‑1000000x** dataset, authored by *ianncity* and released in March 2026, contains one million distilled reasoning traces generated from the KIMI‑K2.5 model on high‑level reasoning tasks. Packaged as a JSON‑line collection of roughly 20 GB (about 5 billion tokens), it falls into the 100 K < n < 1 M size category and is licensed under Apache‑2.0, allowing unrestricted use without attribution.

The dataset is organized into four configurations—General‑Distillation, PHD‑Science, General‑Math, and MultilingualSTEM—each focusing on a different domain. Its distribution spans coding (50 % of entries, covering languages such as Python, C++, Java, and more), science (20 % with a dedicated PHD‑Science subset), mathematics (15 % with a 200 k entry math subset), computer science, logical questions, creative writing, and a multilingual STEM subset. The tags and task categories indicate that the data is suitable for **text generation**, **question answering**, **instruction tuning**, **chain‑of‑thought reasoning**, and **supervised fine‑tuning (SFT)**.

Data were collected using a modified Datagen tool from the TeichAI community over roughly 80 hours. Because the traces are already distilled from a strong reasoning model, they serve as high‑quality examples for training or evaluating large language models on reasoning‑heavy tasks, especially in coding assistance, scientific QA, and multilingual STEM problem solving.