celsowm/jurisprudencias_br

Brazilian jurisprudence (jurisprudência) collected by Juriscraper, published as Parquet for direct consumption by the Hugging Face Dataset Viewer, datasets, DuckDB, Polars, Pandas, and PyArrow.

The default configuration is a Viewer-friendly projection with metadata and ementas. The complete payload is preserved in this same dataset under full/ for bulk downloads and analysis; it is intentionally not registered as a Viewer configuration because it is too large for Dataset Server indexing.

Current dataset coverage

This published snapshot contains decisions for the following values in the tribunal column:

• stf
• stj
• tjrj

The Juriscraper repository also implements collectors for STF, STJ, TJRJ, TJSP, TJPA, TST, Justiça do Trabalho (TST, CSJT and TRT1–TRT24), and CJF. Those sources appear in this dataset only after records have been collected and included in a publication; collector availability does not mean that this snapshot already contains their decisions.

Collection timestamp semantics

• scraped_at: ISO-8601 UTC timestamp recorded when Juriscraper stored the row in its local operational SQLite database (collection time, not the court's publication time).
• data_julgamento / data_publicacao: normalized YYYY-MM-DD strings derived from the court payload when available; may be null.

Public schema

Default configuration

The default configuration intentionally omits the two largest payload columns (inteiro_teor and data_json) so the Dataset Viewer can browse the corpus. Use hf://datasets/celsowm/jurisprudencias_br/full/*.parquet when the complete payload is needed.

The local autoincrement id is not published. The stable logical document identity is the pair (tribunal, unique_key), which is also the deduplication key used by the scraper.

Default schema version: 2

Complete dataset

The Parquet files under full/ use the complete schema below. Their schema version is 1; their data_json column preserves the complete original source item so court-specific fields not in the flat columns remain available.

Process parties (polos)

partes is a UTF-8 JSON array of the process parties (polos do processo), one object per party:

[
  {"nome": "MUNICÍPIO DE RIO DAS OSTRAS", "polo": "ativo",  "tipo": "APELANTE"},
  {"nome": "EDUARDO COIM MARTIN",         "polo": "passivo", "tipo": "APELADO"}
]

• nome: party name (preserved with original capitalization/accents).
• polo: canonical side — ativo (plaintiff/appellant) or passivo (defendant/appellee). Derived from tipo when the source exposes a label.
• tipo: the raw role label as the court presents it (e.g. APELANTE, APELADO, AGTE, AGDO, AUTOR, RÉU, RECORRENTE, RECLAMANTE...). Sources without structured parties (e.g. TJRJ) are enriched from the process detail page; some courts may leave partes null.

Source-specific variability

Different courts return different payloads. Fields outside the flat schema are embedded verbatim in data_json (UTF-8 JSON). Always read data_json when you need fields not listed above.

Licensing / provenance

Decisions are public-domain government works in Brazil, but verify before any commercial use. Provenance is the respective court website; Juriscraper is a non-official collector.

Update cadence

Published as full deterministic snapshots from the authoritative local SQLite database. See the project's docs/HUGGINGFACE_DATASET_PUBLISHING_REFACTOR_PLAN.md.

Usage

With datasets

from datasets import load_dataset

ds = load_dataset("celsowm/jurisprudencias_br")

With DuckDB / SQL

SELECT tribunal, count(*)
FROM 'hf://datasets/celsowm/jurisprudencias_br/data/*.parquet'
GROUP BY tribunal;

With PyArrow / Pandas

import pyarrow.dataset as ds
dataset = ds.dataset(
    f"hf://datasets/celsowm/jurisprudencias_br/data/",
    format="parquet",
)

Snapshot metadata

• Snapshot boundary (max local id): 1029008
• Published rows: 1001292
• Shards: 21

Legacy raw snapshot

If present, jurisprudencia.db is a raw operational snapshot of the SQLite storage engine used during scraping. It is not the preferred ML format; prefer the Parquet files under data/ for analysis and training.