Open-Thoughts-114k: 114,000 Synthetic Reasoning Examples for LLMs

Open-Thoughts-114k contains 114,000 synthetic reasoning examples across math, science, code, and puzzles, released by the open-thoughts team in 2025. This dataset served as the primary training data for the OpenThinker-7B and OpenThinker-32B models. It provides structured thought-process chains designed to improve logical deduction in language models.

Use Cases

• Fine-tuning models for step-by-step mathematical reasoning using the math-focused examples
• Improving programming logic and code generation via the included code reasoning tasks
• Training models on complex logical puzzles to enhance general problem-solving capabilities

Strengths

• 114,000 high-quality reasoning examples
• Multi-domain coverage including math, science, code, and puzzles
• Apache 2.0 license for open redistribution and use

Limitations

• Synthetic generation may result in hallucinated facts or logical errors within the reasoning chains
• Lack of human-verified ground truth for the intermediate reasoning steps

Provenance

Source

open-thoughts (Arxiv: 2506.04178)

Collection Method

synthetic

Freshness

Last updated August 2025