ufca-llms/normas-tcu

NormasTCU is a Brazilian Portuguese legal IR test collection composed of normative documents from the Brazilian Federal Court of Accounts (TCU). These normative acts may have internal effects (e.g., rules governing internal procedures) or external effects (e.g., rules regulating how the court interacts with other public institutions) and differ from jurisprudential documents in both purpose and structure. Jurisprudential documents typically describe specific cases and present the legal reasoning that leads to a decision, whereas normative acts establish general rules that must be applied to specific cases. Moreover, they exhibit a distinct textual structure: jurisprudential documents are usually organized as narrative legal arguments, while normative acts are structured hierarchically into articles, paragraphs, and other legal provisions that define prescriptive rules.

In its original form, NormasTCU contains 14,469 legal documents. Each document corresponds to a normative act and is structured into multiple metadata and textual fields. Among these, the most relevant are TEXTONORMA, which contains the full text of the normative act, and ASSUNTO, a summary describing its content. The ASSUNTO field can help users and retrieval systems identify documents relevant to a given query. However, this field is missing in more than 7,000 documents.

The collection includes 46 queries, some expressed as short keyword queries (less informative and closer to real user searches) and others written in a more semantically rich form. Relevance judgments were produced by four domain experts using a three-level graded scale. In total, the qrels contain assessments for 812 unique judged query–document pairs.

For the JUÁ benchmark, we represent each document by concatenating the ASSUNTO and TEXTONORMA fields, combining the norm summary with its full text. The dataset is publicly available on Hugging Face.

Rank	Model	NormasTCU
1	bm25/anserini	0.3636
2	ufca-llms/Qwen3-Embedding-4B-jua-jurisprudencia (rerank)	0.3083
3	Qwen/Qwen3-Embedding-4B	0.3070
4	ufca-llms/Qwen3-Embedding-4B-jua-jurisprudencia	0.3054
5	Qwen/Qwen3-Embedding-4B (rerank)	0.3022
6	Qwen/Qwen3-Embedding-8B (rerank)	0.2855
7	Qwen/Qwen3-Embedding-8B	0.2798
8	text-embedding-3-small (rerank)	0.2675
9	text-embedding-3-small	0.2625
10	ufca-llms/Qwen3-Embedding-0.6B-jua-v2 (rerank)	0.1940
11	ufca-llms/Qwen3-Embedding-0.6B-jua-v2	0.1748
12	tencent/KaLM-Embedding-Gemma3-12B-2511	0.1087
13	random/384-42	-