Sonar Municipal: TREC-style benchmark and 241k action corpus for semantic search over Brazilian municipal bills

Companion data for the undergraduate thesis "Mineração de Dados e Busca Semântica Aplicadas à Análise de Projetos de Lei Municipais" (Sonar Municipal project, ICMC-USP). Contains a TREC-style retrieval benchmark with 50 topics, 3,792 graded relevance judgments by three LLM judges (Codex/GPT-5.4, Claude Sonnet 4.6, Gemini 3 Pro), three system runs (BM25 over ementas, E5 over ementas, E5 over action-form textualizations), a national corpus of 241,111 ementa→action pairs from public Brazilian municipal SAPL portals, and the supervised training set (999 pairs) used to fine-tune the companion PTT5 base model.