KR-Housing-LongRAG-Bench

A copyright-safe Korean long-context benchmark for evaluating long-context LLMs, RAG systems, and table/tool pipelines over real housing announcements, public tabular data, and housing statutes. The public release contains QA labels, evidence locators, deterministic predicates, answerability labels, split and provider/region metadata, and long-context-bundle references. It does not redistribute raw PDF/HWP/HWPX documents, bundle text, API keys, or hidden gold answers. v0.8 is a human-review repair build (1,997 QA): all positional-cloze probes were regenerated into natural source-grounded questions or removed, and location/answer errors were fixed (LLM-assisted gpt-5.4 + Claude cross-model review).