Gemini 3 Pro High-Reasoning Traces for Model Distillation

250 reasoning traces were generated using the Gemini 3 Pro Preview model with its reasoning depth set to high. The dataset was created by TeichAI in December 2025 to facilitate the distillation of Gemini 3 Pro's reasoning capabilities into open-source large language models. It builds upon a previous 250-example collection.

Use Cases

• Fine-tuning open-source LLMs on summarized reasoning traces to create distilled versions of Gemini 3 Pro.
• Analyzing the structure and patterns of high-depth reasoning traces generated by a state-of-the-art model.
• Training a student model to predict intermediate reasoning steps given a question and a final answer.
• Benchmarking the reasoning fidelity of distilled models against the original Gemini 3 Pro traces.

Strengths

• Contains 250 high-depth reasoning examples generated by Gemini 3 Pro Preview.
• Builds upon a curated set of questions from a known source (reedmayhew).

Limitations

• Small sample size of 250 examples limits statistical power for training robust models.
• Contains generated data which may propagate any biases or errors present in the source model.
• Lacks metadata on question domains, difficulty, or verification of trace correctness.

Provenance

Source

TeichAI via Hugging Face.

Collection Method

Generated using the Gemini 3 Pro Preview API with high reasoning depth setting; some questions sourced from reedmayhew.

Freshness

Last updated December 2025.