ComBench is an Olympiad-level combinatorics benchmark for evaluating rigorous
proof reasoning and constructive realization in large language models. It
contains 100 human-annotated competition-level problems:
50 analysis-centric records in data/analysis.jsonl
50 construction-centric records in data/construction.jsonl
Construction-centric records additionally include explicit construction
instructions, reference constructions, and deterministic Python verifiers.… See the full description on the dataset page:
https://huggingface.co/datasets/Simplified-Reasoning/ComBench.
