Datasets:
language:
- pt
multilinguality: monolingual
task_categories:
- text-retrieval
task_ids: []
config_names:
- corpus
tags:
- text
dataset_info:
- config_name: default
features:
- name: query-id
dtype: string
- name: corpus-id
dtype: string
- name: score
dtype: float64
splits:
- name: test
num_examples: 8205
- config_name: corpus
features:
- name: _id
dtype: string
- name: title
dtype: string
- name: text
dtype: string
splits:
- name: corpus
num_examples: 105669
- config_name: queries
features:
- name: _id
dtype: string
- name: text
dtype: string
splits:
- name: queries
num_examples: 692
configs:
- config_name: default
data_files:
- split: test
path: qrels/test.jsonl
- config_name: corpus
data_files:
- split: corpus
path: corpus.jsonl
- config_name: queries
data_files:
- split: queries
path: queries.jsonl
Ulysses-RFCorpus is a Brazilian Portuguese corpus for legislative LIR with explicit relevance feedback collected in a real production scenario at the Brazilian Chamber of Deputies. The corpus was designed to capture user feedback from the institution's own retrieval workflow, rather than synthetic labels.
The Brazilian Chamber of Deputies includes a specialized department, Legislative Consulting, which supports parliamentarians during the law-making process. In this workflow, parliamentarians can request this department to draft new bills. The queries in Ulysses-RFCorpus correspond to these requests, anonymized, and reflect real information needs. The document collection is composed of legislative proposals (bills). The relevance judgments were provided by 54 legislative consultants using a three-level scale over a pool of retrieved documents, and they could also manually indicate additional relevant bills.
According to the authors, the resource is especially relevant because publicly available legal corpora with expert or user-provided relevance feedback are still scarce, particularly in the legislative domain. The collection includes 692 queries with relevance judgments, reflecting diverse legislative information needs and real user feedback patterns captured during operational retrieval sessions. In this sense, Ulysses-RFCorpus complements predominantly judicial benchmarks by providing supervision aligned with legislative drafting and policy-support tasks.
Within JUÁ, this subset brings relevance judgments grounded in the real operational needs of legislative consultants. This improves ecological validity and helps evaluate whether models generalize beyond lexical matching to the kind of judgments observed in practical legal-government use. The adapted dataset is publicly available on Hugging Face.
| Rank | Model | UlyssesRF |
|---|---|---|
| 1 | bm25/anserini | 0.5313 |
| 2 | Qwen/Qwen3-Embedding-4B (rerank) | 0.4648 |
| 3 | ufca-llms/Qwen3-Embedding-4B-jua-jurisprudencia (rerank) | 0.4604 |
| 4 | Qwen/Qwen3-Embedding-8B (rerank) | 0.4566 |
| 5 | Qwen/Qwen3-Embedding-4B | 0.4505 |
| 6 | ufca-llms/Qwen3-Embedding-4B-jua-jurisprudencia | 0.4407 |
| 7 | text-embedding-3-small (rerank) | 0.4382 |
| 8 | Qwen/Qwen3-Embedding-8B | 0.4355 |
| 9 | text-embedding-3-small | 0.4107 |
| 10 | ufca-llms/Qwen3-Embedding-0.6B-jua-v2 (rerank) | 0.3361 |
| 11 | ufca-llms/Qwen3-Embedding-0.6B-jua-v2 | 0.2791 |
| 12 | random/384-42 | 0.0003 |
| 13 | tencent/KaLM-Embedding-Gemma3-12B-2511 | - |