mquake-wikidata-construction-params

IN premise — summaries/2026/08/24/zhong-2023-mquake-s2-problem-definition.md

Created 2026-08-25T02:59:05+00:00

MQuAKE chains are sampled from a Wikidata subgraph restricted to 37 manually selected relations and top 20% common entities by Wikipedia hyperlink count, with chains of 2-4 triples

Summary

The reasoning tasks in this system are built from a deliberately narrowed slice of Wikidata, using only 37 hand-picked relationship types and the most Wikipedia-linked entities, and each question chain spans just 2 to 4 connected facts. This means the system will perform well on questions about well-known subjects and common relationship patterns, but it will not generalize to obscure entities, long reasoning paths, or relationship types that were excluded during curation.