lucas-lage/SB2-Dataset

SB² Dataset: Similar Brazilian State Bills

The SB² Dataset is a dataset of Brazilian legislative bills annotated with institutional similarity links established during the legislative workflow. It was created to support research on Legal Information Retrieval (LIR), Semantic Textual Similarity (STS), Retrieval-Augmented Generation (RAG), document retrieval, and legislative recommendation systems.

The dataset accompanies the paper:

ReSB²: Machine-Assisted Linking of Legislative Bills using Domain-Adapted ModernBERT and Explainable AI
ACM Hypertext 2026

Dataset Description

The dataset combines legislative bills from two Brazilian legislative bodies:

• Legislative Assembly of Minas Gerais (ALMG)
• Brazilian Chamber of Deputies

Unlike existing Brazilian legislative datasets, SB² includes expert-validated similarity relationships between legislative bills. These links were established by legislative analysts during the regular legislative process and represent institutional judgments regarding substantive legislative relatedness.

The dataset can be used for:

• Continued pre-training using Masked Language Modeling (MLM)
• Semantic Textual Similarity (STS)
• Dense retrieval
• Bi-Encoder training
• Cross-Encoder training
• Legal Information Retrieval
• Retrieval-Augmented Generation (RAG)
• Legislative recommendation systems

Repository Structure

data/
├── raw/
│   ├── almg/
│   │   ├── detalheProposicao.json.gz
│   │   ├── sessoes-legislativas.json
│   │   └── textoProposicao2.json.gz
│   │
│   └── chamber_of_deputies/
│       └── dataset_proposicoes_20240416.csv
│
└── preprocessed/
    ├── almg/
    │   ├── sb2_almg_mlm.csv
    │   └── sb2_almg_pairs.csv
    │
    └── chamber_of_deputies/
        ├── sb2_chamber_mlm.csv
        └── sb2_chamber_pairs.csv

Raw Data

The raw directory contains the original legislative data collected from the official sources.

Preprocessed Data

The preprocessed directory contains the datasets generated after the preprocessing pipeline described in the accompanying paper.

The scripts used to transform the raw legislative data into the preprocessed datasets are publicly available in the ReSB² Framework repository:

https://github.com/LucasLage/ReSB2-Framework

MLM Files

The files

• sb2_almg_mlm.csv
• sb2_chamber_mlm.csv

are intended for continued pre-training using the Masked Language Modeling (MLM) objective.

Each file contains the following columns:

Column	Description
doc_id	Unique legislative bill identifier.
preprocessed_text	Preprocessed legislative bill text.

Similarity Pair Files

The files

• sb2_almg_pairs.csv
• sb2_chamber_pairs.csv

contain document pairs for supervised similarity learning (e.g., Bi-Encoder and Cross-Encoder training).

Each file contains the following columns:

Column	Description
doc_id_1	Identifier of the first bill.
preprocessed_text_1	Preprocessed text of the first bill.
doc_id_2	Identifier of the second bill.
preprocessed_text_2	Preprocessed text of the second bill.
label	Similarity label (1 = similar, 0 = not similar).

Similarity Labels

Positive pairs correspond to similarity links established by legislative analysts during the regular legislative workflow.

These labels represent institutional legislative similarity, rather than purely lexical or semantic similarity, and therefore incorporate domain knowledge and procedural legislative criteria.

Statistics

Source	Bills	Similar Pairs
Legislative Assembly of Minas Gerais (ALMG)	49,868	3,754
Brazilian Chamber of Deputies	71,898	43,484

Citation

If you use this dataset, please cite:

@inproceedings{lage2026resb2,
  title={ReSB²: Machine-Assisted Linking of Legislative Bills using Domain-Adapted ModernBERT and Explainable AI},
  author={Lage, Lucas Gabriel and others},
  booktitle={Proceedings of the ACM Conference on Hypertext and Social Media},
  year={2026}
}

License

Please refer to the licenses of the original legislative data providers.

This repository distributes processed versions of publicly available legislative documents for research purposes.

Acknowledgments

This work was supported by the Legislative Assembly of Minas Gerais (ALMG), CNPq, CAPES, and FAPEMIG.