GuiiCorreia/ulysses-rfcorpus-mtrag

Ulysses-RFCorpus — unmodified mirror used by Ulysses-MTRAG

This dataset mirrors, unmodified, the two files of Ulysses-RFCorpus that the Ulysses-MTRAG benchmark (IEEE ICTAI 2026) is built on: the collection of 105,669 legislative bills of the Brazilian Chamber of Deputies and the 692 expert queries with graded relevance labels produced by the Chamber's legislative consultants.

Provenance and citation

Vitório, D.; Souza, E.; Martins, L.; da Silva, N. F. F.; de Carvalho, A. C. P. L. F.; Oliveira, A. L. I.; de Andrade, F. E. "Building a relevance feedback corpus for legal information retrieval in the real-case scenario of the Brazilian Chamber of Deputies." Language Resources and Evaluation 59:1257–1277, 2025. doi:10.1007/s10579-024-09767-3

The material is redistributed with attribution as authorised by one of its authors. Please cite the paper above when using it.

@article{vitorio2025ulysses,
  title   = {Building a relevance feedback corpus for legal information retrieval in the real-case
             scenario of the {B}razilian {C}hamber of {D}eputies},
  author  = {Vit{\'o}rio, Douglas and Souza, Ellen and Martins, Lucas and da Silva, N{\'a}dia F. F.
             and de Carvalho, Andr{\'e} C. P. L. F. and Oliveira, Adriano L. I. and de Andrade, Fl{\'a}vio E.},
  journal = {Language Resources and Evaluation},
  volume  = {59},
  pages   = {1257--1277},
  year    = {2025},
  doi     = {10.1007/s10579-024-09767-3}
}

Files

Both files are byte-identical to the release used in the experiments; nothing was filtered, renamed or reordered.

Schema

bills_dataset.csv (one row per bill): code (int), sig_tipo (str, e.g. PL), name (str, e.g. PL 3650/2021 — the identifier used as join key), txt_ementa (str, official summary), em_tramitacao, situacao (str), text (str, full text), text_preprocessed (str). Years covered: 1992–2022.

relevance_feedback_dataset.csv (one row per expert query): id (int), query (str), user_feedback (JSON list of {"id": "<bill name>", "class": "r"|"pr"|"i", "score", "score_normalized"}; r = relevant, pr = partially relevant, i = irrelevant), extra_results (JSON list of additional relevant bill ids), date_created, num_doc_feedback (int), extra_results_size (int).

Used by

• —Ulysses-MTRAG — code, benchmark, judgments and results: https://github.com/GuiiCorreia/ulysses-mtrag-pt
• —Paper: Ulysses-MTRAG: A Synthetic Multi-Turn Benchmark for Retrieval-Augmented Generation over Brazilian Legal Text, IEEE ICTAI 2026 (Dutra, Caraíba, Silva, Santos, Teles, Soares, Silva — INF/UFG).

Licence

The corpus material in this dataset is subject to the terms of the original Ulysses-RFCorpus release (see the paper above); it is mirrored here with attribution as authorised by one of its authors. Nothing in this dataset is authored by the Ulysses-MTRAG team; our own material (generated turns, labels, judgments, code) lives in the GitHub repository under CC BY 4.0 / MIT.