Most math datasets focus on lower-level problems or small expert-curated benchmarks; ResearchMath-14k instead scales research-grade, open-problem statements by combining automated extraction with agentic refinement. That makes it one of the largest corpora of self-contained research problems intended explicitly for model understanding, prompt construction, and fine-tuning toward research-level mathematical reasoning.
What Sets It Apart
- Scale + research focus: 14,056 items spanning 11 top-level mathematical domains (Analysis/PDEs, Mathematical Physics, Combinatorics, Geometry/Topology, Algebra, etc.), giving broader coverage than typical evaluation suites. This breadth helps train models to recognize domain-specific problem structures rather than only elementary manipulations.
- Agentic two-stage pipeline: an extractor agent finds candidate open questions in source documents and a refiner agent verifies status, assigns taxonomy labels, and rewrites each item into a self-contained problem statement. The pipeline trades manual curation for scalable, repeatable extraction while preserving metadata about source and open/solved status.
- Explicit open-status and taxonomy metadata: each record includes open/partially_solved/solved/unknown tags (about 59% marked open) plus hierarchical taxonomy labels, enabling filtered splits for training (e.g., only open problems) or targeted evaluation (domain-specific difficulty slices).
Intended Use and Practical Tradeoffs
Great fit if you want large-scale, domain-diverse supervision or evaluation data for models that must handle research-grade problem statements — for tasks such as question understanding, prompt engineering for long-form reasoning, or fine-tuning on partially specified problems. The dataset’s metadata supports experiments that differentiate open problems from known results or that focus on particular subfields.
Look elsewhere if you require ground-truth step-by-step proofs or verified solutions: many records are labeled open or partially solved and lack complete canonical solutions, so the corpus is better for problem understanding, formulation, and generating attempted solutions than for supervised solution verification. Also, because items are extracted from heterogeneous sources, model authors should expect variability in notation, background assumptions, and required prerequisites and consider preprocessing or filtering for consistent training signals.
Where It Fits
ResearchMath-14k occupies the gap between very large, lower-level math corpora and small, expert-curated research benchmarks. Use it to scale data-driven attempts at research-level mathematical reasoning, to build prompt templates for open problems, or to create evaluation sets that stress cross-domain problem recognition rather than only arithmetic skill.
