MHGB: A Multi-Hop Knowledge Graph Benchmark for Legal Reasoning Evaluation

MHGB is a benchmark for evaluating multi-hop legal reasoning over Russian statutory law. It represents statutory knowledge as a typed graph of 5,519 article nodes connected by 8,760 relations of five types across ten federal codes, and uses that structure to generate 600 multi-hop reasoning tasks, each carrying a gold reasoning chain. Every task is posed in two modes: closed-book, answered from parametric knowledge alone, and open-book, with graph-derived statutory context. Responses are scored at three levels by two LLM judges held outside the leaderboard, and the GAP metric measures the change between the two modes. This deposit contains the full knowledge graph, the statutory corpus, a stratified 120-task public subset with gold reasoning chains, per-model scores for twelve evaluated models, the complete evaluation pipeline and an interactive graph explorer. Code is licensed under Apache-2.0, derived data under CC BY-4.0; the statutory texts are in the public domain.