High-quality, verifiable math reasoning traces are a bottleneck for supervised fine-tuning of LLMs; this dataset supplies both long-form chain-of-thought and Python tool-integrated solution trajectories where final answers are checked against verified references.
What Sets It Apart
- Scale and modalities: 545,431 samples split into 285,516 chain-of-thought (COT) traces and 259,915 tool-integrated (TIR) traces, totaling over 6.3 billion tokens.
- Verified answers: only solutions whose final answers match reference answers are retained, reducing noisy labels for supervised training and evaluation.
- Practical format and size: stored as 12 zstd-compressed Parquet shards (≈5.16 GiB on disk, ≈18.43 GiB uncompressed JSONL), ready for dataset libraries and pipeline ingestion.
- Provenance & licensing: seed problems drawn from AoPS and Math StackExchange/MathOverflow; AoPS-derived samples are CC BY 4.0 while Math.SE-derived samples use CC BY‑SA 4.0, tracked per sample.
- Trace generation pipeline: solution trajectories produced by DeepSeek‑V4‑Pro (High inference mode) and generated via NeMo‑Skills tooling, enabling reproducible multi-mode supervision.
Who It's For and Tradeoffs
Great fit if you need a large, curated supervised corpus to fine-tune or evaluate LLMs on structured mathematical reasoning (including experiments comparing COT vs. tool-assisted traces, long-context training, or answer-verification pipelines). Look elsewhere if you require human-authored proof texts (proof-style problems were filtered) or need an exclusively permissive-license corpus (some samples are CC BY‑SA). Also note model-generated traces can reflect generator biases; use verification and validation in downstream evaluation.
