Gemini 3 Pro Preview High Reasoning 1000X: Synthetic Q&A for Model Distillation

A reasoning dataset containing 1000 examples generated using the Gemini 3 Pro Preview API with its reasoning depth set to high. The dataset was created by author kabsis and last updated on December 19, 2025. It is intended for fine-tuning open-source language models on summarized reasoning traces.

Use Cases

• Fine-tuning open-source LLMs based on high-depth reasoning traces mentioned in the description
• Conducting knowledge distillation experiments based on the synthetic question-and-answer pairs
• Benchmarking model reasoning capabilities based on the generated high-reasoning questions
• Creating instruction-following datasets for alignment training based on the API-generated content

Strengths

• Contains 1000 reasoning examples, as indicated by the title.
• Generated using a high-reasoning-depth setting of the Gemini 3 Pro Preview API.
• Includes 250 examples sourced from an existing dataset (TeichAI/gemini-3-pro-preview-high-reasoning-250x).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but the specific composition and balance of question types are not detailed.
• Data may reflect source bias inherent to the generating model and its training data.

Provenance

Source

huggingface

Collection Method

Generated via the Gemini 3 Pro Preview API, with some questions sourced from 'reedmayhew' and others newly generated.

Time Range

Creation date is not specified; last update was 2025-12-19.

Freshness

Last updated 2025-12-19 09:48:11; freshness should be verified.

Geography

Spatial coverage is not specified.