OpenThoughts-114k Synthetic Reasoning Dataset

## Overview OpenThoughts-114k is an open synthetic reasoning dataset containing approximately 114,000 high-quality reasoning examples spanning math, science, code, and puzzles. It was used to fine-tune the OpenThinker-7B and OpenThinker-32B models. The data is distributed in parquet format and is text-modality. Released by the Open Thoughts project in 2025, with an accompanying paper (arXiv:2506.04178). ## Schema - `system` — string — system prompt used during generation - `conversations` — list — multi-turn conversation with user prompt and assistant reasoning trace - Additional metadata fields per subset (problem domain, source, verification info) — see HF dataset card for the full schema across subsets The `default` subset is the ready-to-train SFT format used for the OpenThinker models. ## Sources - Open Thoughts project — https://huggingface.co/datasets/open-thoughts/OpenThoughts-114k — Apache 2.0 - Reference paper: https://arxiv.org/abs/2506.04178 ## Methodology Reasoning traces were generated synthetically by prompting strong reasoning models on curated seed problems across math, science, code, and puzzle domains, then filtered/verified for quality. The pipeline was built with the Curator framework. The `default` subset is the deduplicated, formatted SFT-ready slice that produced the OpenThinker checkpoints; other subsets retain richer per-example metadata for inspection and reweighting. ## Known gaps & limitations As a synthetic distillation dataset, traces reflect the biases and error modes of the teacher model(s) used to generate them. Coverage is weighted toward math and code; humanities and open-ended reasoning are under-represented. The dataset is not deduplicated against popular reasoning benchmarks (MATH, GSM8K, HumanEval, etc.) — leakage risk if used to train models intended for evaluation on those suites. English-dominant. ## Intended use & out-of-scope - IS for: SFT/fine-tuning reasoning models, distillation research, building chain-of-thought datasets, RAG over solved-problem corpora, eval prompt mining. - NOT for: training models you intend to evaluate on MATH/GSM8K/HumanEval/LiveCodeBench without first running contamination checks; production deployment of generated answers without verification. _Federated dataset: 18 parquet shards, 3.30 GB total. Queries and downloads stream through the DataBazaar API._ ## Temporal validity This dataset includes column(s) keyed on recycled identifiers — the same value can refer to different entities at different times: - **domain name** (reissued by registrars (drop-catching)) — domains are recycled; use WHOIS history to filter enrichment by the current registration interval. Original supplier listing: OpenThoughts-114k: Synthetic Reasoning Dataset 114k high-quality synthetic reasoning examples across math, science, code, and puzzles. Used to fine-tune OpenThinker-7B/32B models. Apache 2.0 licensed, parquet format.