thiagoambiel/sonar-municipal-pl-actions

Sonar Municipal — PL Actions Corpus

The largest publicly released Portuguese legal text rewriting dataset: 241,111 pairs mapping the original ementa (summary) of a Brazilian municipal Projeto de Lei (PL) to its action-form textualization (ação). Action form is a direct, imperative rewrite that strips juridical boilerplate, normalizes typography, and surfaces the underlying intervention rather than the legal instrument.

Companion to: ICMC-USP undergraduate thesis "Mineração de Dados e Busca Semântica Aplicadas à Análise de Projetos de Lei Municipais" (Sonar Municipal project). The full dataset record (with the TREC-style retrieval benchmark, annotation pool, and control experiments) lives on Zenodo: DOI 10.5281/zenodo.20564283.

Configurations

This dataset has two configurations:

default — full corpus (241,111 rows)

Inference outputs from the thiagoambiel/sonar_municipal_ptt5_ementa2action model over the national corpus of Brazilian municipal PLs scraped from public SAPL portals. Each row carries the original bill metadata (municipality, state, presentation date, public URL, SAPL ids, etc.) alongside the ementa/acao pair — see Schema.

from datasets import load_dataset
ds = load_dataset("thiagoambiel/sonar-municipal-pl-actions")
print(ds["train"][0])
# {'ementa': 'DISPÕE SOBRE A IMPLANTAÇÃO DE ESTUFAS COM HORTAS PRODUZIDAS COM GARRAFAS PET...',
#  'acao': 'Instalar estufas com hortas produzidas com garrafas pet nas escolas municipais.',
#  'municipio': 'Marabá', 'uf': 'PA', 'materia_id': 27075, 'numero': 1, 'ano': 2025,
#  'data_apresentacao': '2025-03-14', 'tipo_label': 'ANT Anteprojeto de Lei ', 'tipo_id': 9,
#  'em_tramitacao': True,
#  'link_publico': 'https://sapl.maraba.pa.leg.br/materia/27075/acompanhar-materia/',
#  'sapl_url': 'https://sapl.maraba.pa.leg.br',
#  'doc_id': 'https://sapl.maraba.pa.leg.br::materia::27075'}

supervised — training set (999 rows)

The curated supervised set used to fine-tune the companion model. Labels initially generated by GPT-5.1-Thinking, with manual quality control of all pairs by the thesis author (error taxonomy: polarity reversals, fantasy-name preservation, object omission, juridical sense shifts). This is the right config if you want to train your own seq2seq on this task.

ds = load_dataset("thiagoambiel/sonar-municipal-pl-actions", name="supervised")
print(len(ds["train"]))   # 999

To reproduce the original train/val split (deterministic, seed 42):

import random
rows = list(ds["train"])
random.Random(42).shuffle(rows)
train, val = rows[:900], rows[900:]   # 900 / 99 split

Schema

The default config has 14 columns: the ementa/acao text pair plus the original bill metadata. (The supervised config keeps only ementa/acao.)

{
  "ementa": "DISPÕE SOBRE A IMPLANTAÇÃO DE ESTUFAS COM HORTAS PRODUZIDAS COM GARRAFAS PET NAS ESCOLAS MUNICIPAIS DE MARABÁ E DA OUTRAS PROVIDÊNCIAS.",
  "acao": "Instalar estufas com hortas produzidas com garrafas pet nas escolas municipais.",
  "municipio": "Marabá",
  "uf": "PA",
  "materia_id": 27075,
  "numero": 1,
  "ano": 2025,
  "data_apresentacao": "2025-03-14",
  "tipo_label": "ANT Anteprojeto de Lei ",
  "tipo_id": 9,
  "em_tramitacao": true,
  "link_publico": "https://sapl.maraba.pa.leg.br/materia/27075/acompanhar-materia/",
  "sapl_url": "https://sapl.maraba.pa.leg.br",
  "doc_id": "https://sapl.maraba.pa.leg.br::materia::27075"
}

Column definitions

Column	Type	Description
ementa	string	Original bill summary (verbatim from SAPL, typography normalized).
acao	string	Action-form textualization (model output).
municipio	string	Municipality name.
uf	string	State (2-letter UF code).
materia_id	int64	Bill (matéria) id within its SAPL instance.
numero	int64	Bill number.
ano	int64	Bill year.
data_apresentacao	string	Presentation date, YYYY-MM-DD (falls back to data de recebimento when the presentation date is absent at the source).
tipo_label	string	Bill-type label from SAPL (verbatim, may carry a trailing space).
tipo_id	int64	Bill-type id within its SAPL instance.
em_tramitacao	bool	Whether the bill is still under processing.
link_publico	string	Public URL of the bill on the municipality's SAPL portal.
sapl_url	string	Base URL of the SAPL instance.
doc_id	string	Stable key "{sapl_url}::materia::{materia_id}"; unique across the corpus and shared with the qrels/recommendations of the TREC benchmark in the Zenodo record.

• Encoding: UTF-8.
• HTML entities normalized (e.g., – → –).
• Newlines collapsed.
• (municipio, uf, materia_id, numero, ano, …) are the source bill metadata; only acao is model-generated.

Behaviors learned

Looking at the corpus, the action-form rewrites consistently:

• Remove juridical boilerplate (E DÁ OUTRAS PROVIDÊNCIAS, Dispõe sobre, Autoriza o Poder Executivo a).

• Normalize typography (ALL-CAPS → sentence case, preserving acronyms like LIBRAS, SUS, TEA, CIPED, Castramóvel).

• Generalize toponyms (MUNICÍPIO DE MARABÁ → município), except where the locale is policy-relevant.

• Compress conditional clauses into direct infinitive verb + object phrases.

• Express the underlying action rather than the legal instrument:

  Ementa	Naive sintactic rewrite	This dataset's action
  "Institui o Projeto Criança Cidadã para emissão de registro de nascimento dentro das maternidades públicas"	"Criar o Projeto Criança Cidadã"	"Emitir registros de nascimento dentro das maternidades públicas"

Provenance

• Source corpus: public SAPL portals (sapl.<municipio>.<uf>.leg.br). Bill summaries (ementas) are redistributed verbatim under Lei 9.610/98 art. 8º II (textos legislativos não são objeto de proteção autoral).
• Bill metadata: the municipio, uf, materia_id, numero, ano, data_apresentacao, tipo_label, tipo_id, em_tramitacao, link_publico, sapl_url and doc_id columns are reattached to each pair via a deterministic positional join back to the original SAPL scrape (the same record the ementa was extracted from), so the metadata always matches the bill the acao describes.
• Action rewrites: produced by the PTT5 base fine-tune, itself trained on the 999 supervised pairs in the supervised configuration.
• Snapshot date: late 2025. Full methodology in the Zenodo record's PROVENANCE.md (DOI 10.5281/zenodo.20564283).

Use cases

• Information Retrieval: index the acao field with a dense encoder for substantially better query-document semantic match than raw ementas. See the TREC-style benchmark in the parent Zenodo record.
• Sequence-to-sequence training: pair-based fine-tuning of T5 family or seq2seq architectures on Portuguese legal-to-operational rewriting (use the supervised config).
• Policy analysis: action form is more amenable to clustering and topic modeling than verbose juridical boilerplate.
• Document classification: use acao as denoised input for domain classifiers.

Limitations

• Uneven coverage across Brazilian states (some UFs over-represented due to higher SAPL adoption).
• Action rewrites are model-generated at full-corpus scale (BERTScore-F1 ≈ 0.849 on held-out validation). Verify outputs before high-stakes use.
• Manual revision of the seed pairs emphasized policy-area diversity; technical subdomains (e.g., urban planning bills with engineering jargon) may be less reliable.
• No PII of private individuals. Where ementas mention public officials (vereadores), this is in their public capacity (LGPD art. 7º III + art. 4º III). No vote records, no contact details, no individual authorship attributions.

Citation

@dataset{ambiel_sonar_municipal_dataset_2026,
  title     = {Sonar Municipal: TREC-style benchmark and 241k action corpus for semantic search over Brazilian municipal bills},
  author    = {Ambiel, Thiago},
  year      = 2026,
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20564283},
  version   = {1.0.0},
  note      = {Supervised by André C. P. L. F. de Carvalho (ICMC-USP). Supported by the AI4PEP network.}
}

ABNT NBR 6023 (PT-BR):

AMBIEL, Thiago. Sonar Municipal: benchmark TREC e corpus de 241 mil ações para busca semântica de Projetos de Lei municipais brasileiros (versão 1.0.0). [Dataset]. Zenodo, 2026. DOI: 10.5281/zenodo.20564283.

Acknowledgements

Supported by the AI4PEP network (Artificial Intelligence for Public Health Emergency Preparedness) — acknowledged in the thesis; no structured grant identifier is associated with this record. Supervised by Prof. André C. P. L. F. de Carvalho (ICMC-USP). Bill data sourced from SAPL/Interlegis (Câmara dos Deputados).

Contact

For data corrections or LGPD takedown requests: thiago.ambiel@usp.br. See the parent Zenodo record's PROVENANCE.md § 11 for the formal procedure.