Reasoning Traces for Code, Science, and Math from KIMI-K2.5

700,000 reasoning traces were distilled from the KIMI-K2.5 language model on high-reasoning tasks. The dataset contains 2.5 billion tokens, with a distribution of 50% coding, 20% science, 15% math, 5% computer science, 5% logical questions, and 5% creative writing. It was created by author 'ianncity' and last updated in April 2026.

Use Cases

• Fine-tune language models on the 'Coding' subset traces to improve performance on tasks for Python, C++, and Java.
• Train models for multi-step reasoning using the 'Science' subset traces covering physics, chemistry, and biology concepts.
• Use the 'Math' subset traces containing algebra and calculus problems to enhance a model's mathematical reasoning capabilities.
• Analyze the structure of 'Logical Questions' traces to benchmark model performance on formal reasoning tasks.
• Leverage the 'Creative Writing' traces to study how reasoning chains are applied in open-ended text generation scenarios.

Strengths

• Contains 700,000 distinct reasoning traces for model training.
• Covers a broad domain mix with 50% coding, 20% science, and 15% math content.
• Totals 2.5 billion tokens of distilled model reasoning.

Limitations

• Specific row count, column names, and data format are not provided.
• Potential class or domain imbalance is indicated by the fixed percentage distribution.
• The raw description notes a modified data collection method, details of which are not fully specified.

Provenance

Source

Distilled from the KIMI-K2.5 language model.

Collection Method

Collected using a modified Datagen method, as noted in the description.

Freshness

Last updated on 2026-04-03.