ufca-llms/jua

JUÁ-Juris is centered on jurisprudence drawn from the curated jurisprudence collection of the Brazilian Federal Court of Accounts (TCU). In this collection, each instance contains an enunciado and an excerto: the enunciado is an abstractive summary of the ruling, while the excerto is the passage from the ruling that supports that summary. In our retrieval setup, the enunciado serves as the query, and the corresponding excerto is treated as the ground-truth positive passage. Within the benchmark, JUÁ-Juris represents one of the two jurisprudence-focused retrieval settings.

In the TCU context, this curated jurisprudence collection is a subset of precedents chosen for legal relevance, such as novel holdings, decisions with substantial collegial discussion, and reiterated understandings with institutional importance. This curation strategy reduces noise and emphasizes decisions that are more useful for precedent-oriented legal search.

The construction of the JUÁ-Juris dataset follows a difficulty-oriented selection strategy based on lexical retrieval (BM25), with supervised train–test partitioning. We start from a tabular jurisprudence base (pipe-separated) containing, among other fields, enunciado and excerto. Preprocessing removes HTML markup, normalizes specific summary patterns (e.g., súmula-style forms), and discards incomplete records or entries without valid textual content.

To estimate retrieval difficulty, BM25 is applied over the full set of enunciados. For each instance $i$, its own enunciado is used as the query, and the rank position of the corresponding document in the BM25 output is computed. This position is used as a difficulty score (BM25_RANK): a lower self-retrieval rank indicates greater estimated difficulty relative to a lexical baseline.

The test-set construction is then performed via stratified sampling over difficulty bins derived from BM25 ranks. Each query is mapped to a difficulty level using logarithmic bins from 1 to 1024, plus a not_found class for cases where the ground-truth document is not among the top-1000 ranked results. The test split corresponds to 10% of the dataset (1,714 examples), sampled across easy, medium, hard, and not_found strata to preserve coverage across difficulty regimes. The remaining 90% forms the training split.

The relevance judgments, also called qrels, are generated using a direct query–document correspondence (query-id == corpus-id) with binary relevance (score = 1), preserving a simple exact positive-pair retrieval protocol. The dataset is publicly available on Hugging Face.

Methodologically, this design avoids bias toward trivial cases while enabling robustness analysis under progressively harder retrieval scenarios. As a consequence, the benchmark becomes more sensitive to semantic gains from embedding-based retrieval and reranking. A limitation is that the difficulty criterion depends on a single inducer (BM25), which may reflect biases in tokenization and lexical matching.

Rank	Model	Jua
1	ufca-llms/Qwen3-Embedding-0.6B-jua-v2 (rerank)	0.3078
2	ufca-llms/Qwen3-Embedding-0.6B-jua-v2	0.3018
3	ufca-llms/Qwen3-Embedding-4B-jua-jurisprudencia	0.2904
4	ufca-llms/Qwen3-Embedding-4B-jua-jurisprudencia (rerank)	0.2903
5	Qwen/Qwen3-Embedding-8B (rerank)	0.2209
6	Qwen/Qwen3-Embedding-8B	0.2172
7	Qwen/Qwen3-Embedding-4B (rerank)	0.2009
8	Qwen/Qwen3-Embedding-4B	0.1989
9	tencent/KaLM-Embedding-Gemma3-12B-2511	0.1538
10	text-embedding-3-small (rerank)	0.1115
11	text-embedding-3-small	0.1027
12	bm25/anserini	0.1026
13	random/384-42	0.0008