Large-scale math reasoning data is one of the clearest levers for improving LLMs' step-by-step problem solving — but quality, provenance, and format matter more than raw size. IFM/Math-Reasoning packages multiple curated and synthetically processed math-focused splits in a Parquet-first layout designed for streaming training pipelines and dataset inspection.
What Sets It Apart
- Multi-subset organization: five named subsets (math-thinking-qwen, math-thinking-oss, math-rewrite, math-dialogue, socratic-math-dialogue) so you can stream or weight each source independently. This makes it easy to mix high-quality reasoning dialogues with broader synthetic generations.
- Parquet + shard design: every subset is published as Parquet shards for efficient columnar access and integration with streaming loaders (datasets, dask, polars), reducing preprocessing overhead for large-scale training runs.
- Provenance & inspection-first: features and per-record provenance can differ by subset; the README and example code encourage inspecting
featuresbefore building a pipeline, which helps avoid silent schema mismatches. - Production scale: published as part of the K2 Horizon series with multi-terabyte shards and billions of rows (large-scale footprint suitable for pretraining/continued-pretraining).
Who it's for and trade-offs
Great fit if you train or evaluate LLMs on mathematical reasoning and want ready-to-stream Parquet shards that separate dialogue, chain-of-thought, and rewrite data. Look elsewhere if you need small, fully human-verified benchmarks or reproducible, citation-linked problem sources only—this collection may include synthetic generations, filtered/cleaned originals, and provenance metadata that still requires user validation and filtering. Users are responsible for auditing for duplicates, factual errors, and safety-related artifacts.
