OpenR1-Math-220k Reasoning Traces

## Overview OpenR1-Math-220k is a large-scale dataset for training and evaluating mathematical reasoning models. It contains approximately 220,000 math problems sourced from NuminaMath 1.5, each paired with two to four chain-of-thought reasoning traces generated by DeepSeek R1. The dataset is distributed in Parquet format and is English-language. Each problem includes at least one reasoning trace with a verified correct answer, making it directly usable for supervised fine-tuning of reasoning models. ## Schema - `problem` — string — the math problem statement (from NuminaMath 1.5) - `solution` — string — reference solution from the source - `answer` — string — the canonical final answer - `problem_type` — string — category/topic of the problem - `question_type` — string — question format (e.g. proof, numeric) - `source` — string — original competition/source within NuminaMath - `uuid` — string — unique problem identifier - `is_reasoning_complete` — list[bool] — per-trace completion flag - `generations` — list[string] — DeepSeek R1 reasoning traces (2-4 per problem) - `correctness_math_verify` — list[bool] — Math Verify correctness flags - `correctness_llama` — list[bool] — Llama-3.3-70B judge correctness (subset) - `finish_reasons` — list[string] — generation termination reasons - `correctness_count` — int — number of correct traces ## Sources - open-r1/OpenR1-Math-220k on Hugging Face — https://huggingface.co/datasets/open-r1/OpenR1-Math-220k — Apache-2.0 - Underlying problems: NuminaMath 1.5 (AI-MO/NuminaMath-1.5) - Reasoning traces generated by DeepSeek R1 ## Methodology Problems were drawn from NuminaMath 1.5, a curated collection of math competition and textbook problems. For each problem, the Open-R1 team prompted DeepSeek R1 to produce two to four full chain-of-thought reasoning traces. Generated traces were then verified for correctness: the majority were checked programmatically with Math Verify (symbolic answer comparison), and roughly 12% required Llama-3.3-70B-Instruct as an LLM judge for cases where symbolic verification was ambiguous. Only problems with at least one trace having a verified correct answer are retained. The dataset ships in two splits (a default split and a more aggressively filtered split). ## Known gaps & limitations English-only; problems are primarily competition-style math and may not generalize to applied/word-problem domains outside that distribution. Reasoning traces reflect DeepSeek R1's style and biases — fine-tuning on them will inherit those patterns. The ~12% of samples judged by Llama-3.3-70B rather than symbolic verification carry higher false-positive risk on correctness. The source does not document overlap with common math eval suites (MATH, GSM8K, AIME), so leakage risk exists for benchmark evaluation — buyers should decontaminate empirically before reporting eval numbers. ## Intended use & out-of-scope - IS for: supervised fine-tuning of reasoning/chain-of-thought models, distillation from R1, RAG over worked math solutions, evaluating reasoning trace quality. - NOT for: training models intended to be evaluated on MATH/AIME/competition benchmarks without first running decontamination — leakage risk via the NuminaMath source. _Federated dataset: 40 parquet shards, 7.86 GB total. Queries and downloads stream through the DataBazaar API._ _PII signals: cc_shape×1 (Luhn-valid: 0) present in the sample. Common in public datasets (papers, logs) but worth knowing before joining with private data._ Original supplier listing: OpenR1-Math-220k: Math Reasoning Traces from DeepSeek R1 220k math problems with verified DeepSeek R1 reasoning traces, sourced from NuminaMath 1.5. Apache-2.0 licensed dataset for training and evaluating mathematical reasoning models.