Datasets:
license: cc-by-4.0
task_categories:
- reinforcement-learning
- text-generation
- tabular-classification
language:
- en
pretty_name: LeanWordBench
size_categories:
- 10K<n<100K
tags:
- benchmark
- formal-verification
- theorem-proving
- algebra
- reinforcement-learning
- word-problems
- lean
configs:
- config_name: default
data_files:
- split: train
path: data/leanwordbench_instances.jsonl
Dataset Card: LeanWordBench
Summary
LeanWordBench is a benchmark suite for verifiable certificate search over algebraic word problems. Instances are grouped into families such as free-group word equality, subgroup membership, word equations, braid rewriting, exact planning, Andrews--Curtis frontier search, knot move traces, Gröbner reduction, and group-ring witness search.
Intended Use
The benchmark is intended for evaluating systems that produce explicit mathematical certificates: rewrite traces, witnesses, substitutions, action sequences, collision candidates, or algebraic objects. Scores should be reported only after exact replay by the family verifier.
Not Intended For
- Treating frontier timeouts as mathematical negative labels.
- Ranking systems by natural-language explanations that do not include accepted certificates.
- Claiming that classical exact procedures solve frontier regimes merely because they audit released closed slices.
Data Sources
Most rows are generated from standard mathematical presentations and exact
verifiers. External sources are referenced in data/manifest.json and include
SMT-LIB/StringFuzz style word-equation resources, AC-Solver/Miller--Schupp
Andrews--Curtis instances, SymbolicData-inspired Gröbner systems, Regina/knot
diagram resources, and Gardam/Soelberg group-ring artifacts where applicable.
Responsible AI Notes
The benchmark contains symbolic mathematical data, not human-subject, personal, biometric, medical, or demographic records. The main risk is scientific overclaiming: unknown frontier rows must remain unknown unless an explicit certificate or obstruction is found.