One Million Reasoning Traces from KIMI-K2.5 Model

KIMI-K2.5-1000000x contains 1,000,000 reasoning traces distilled from the KIMI-K2.5 model across high-reasoning tasks. The dataset was created by author 'ianncity' and was last updated in April 2026. Its distribution includes coding (50%), science (20%), math (15%), computer science (5%), logical questions (5%), and creative writing (5%).

Use Cases

• Training a model to generate Python code solutions from reasoning traces in the coding subset.
• Fine-tuning a model for step-by-step calculus problem-solving using traces from the math subset.
• Analyzing reasoning patterns for physics or chemistry questions within the science subset.
• Developing a model to follow logical deduction chains from the logical questions subset.
• Studying creative writing structure and narrative development from the creative writing traces.

Strengths

• Contains 1,000,000 reasoning traces, providing a large corpus for training.
• Explicitly defined domain distribution with 50% coding, 20% science, and 15% math content.
• Includes specialized subsets like PHD-Science (100k completions) and kimiMath200k (200k completions).

Limitations

• Specific row counts, column names, and data formats are unknown.
• Potential class imbalance is indicated by the fixed percentage distribution across domains.
• The raw description is minimal, with key details requiring navigation to an external page.

Provenance

Source

Distilled from the KIMI-K2.5 model by author 'ianncity'.

Collection Method

Reasoning traces were distilled from the model on high-reasoning questions.

Freshness

Last updated on 2026-04-07.