Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 1154498677 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

JuDDGES-pl

JuDDGES-pl is a unified Polish court judgment corpus for civil-law legal NLP, accompanying the NeurIPS 2026 Evaluations & Datasets Track submission "JuDDGES-pl: Hundreds of Thousands of Structurally Enriched Polish Judgments from Common and Administrative Courts".

The corpus combines two disjoint branches of the Polish judiciary into a single Hugging Face dataset with two configs:

Config Source Coverage Rows
pl-court JuDDGES/pl-court-raw-enriched Common courts (Supreme Court, courts of appeal, regional and district courts) 100K–1M
pl-nsa JuDDGES/pl-nsa-enriched Administrative courts (Supreme Administrative Court and lower) 1M–10M

Each judgment includes:

  • Verbatim source content (text, signature, court name, judges, dates, statute references) preserved from the official Polish court publication portals.
  • LLM-extracted analytical fields generated by a uniform Google Gemini 2.5 Pro extraction pipeline: factual_state, legal_state, extracted_summary, extracted_thesis, extracted_legal_bases, extracted_keywords, extracted_title.

⚠️ Silver-label commitment

The LLM-extracted fields are silver labels: produced by a single model under uniform prompting, without per-document human validation at corpus scale.

  • Use them for: descriptive analysis, model pretraining, weak supervision, retrieval indexing.
  • Do not use them for: benchmarking with held-out test sets, regulatory or individual legal decisions.
  • Always cite the model version (Google Gemini 2.5 Pro) when reporting analyses derived from these fields, since LLM behavior shifts across releases.

A separate validated subset (JuDDGES/pl-swiss-franc-loans) provides gold annotations for tasks requiring high-fidelity labels.

Loading

from datasets import load_dataset

common = load_dataset("JuDDGES/juddges-pl", "pl-court", split="train", streaming=True)
nsa    = load_dataset("JuDDGES/juddges-pl", "pl-nsa",   split="train", streaming=True)

for row in common.take(3):
    print(row["court_name"], row["signature"], row["extracted_summary"][:120])

Licensing

  • Source judgments: public-domain legal acts under Polish copyright law (Art. 4 pkt 2 of the Ustawa o prawie autorskim — orzeczenia sądów are dokumenty urzędowe and not subject to copyright).
  • Enrichment layer (LLM-extracted analytical fields, structured metadata, Croissant): released under CC BY 4.0.

Privacy

Source judgments are pseudonymized at the court level before publication (personal names of natural persons → initials/codes; PESEL, addresses, dates of birth redacted). Names of public officials, judges, and corporate parties are typically retained as legally permitted. We perform no additional de-anonymization. Users redistributing derivatives must comply with GDPR/RODO and must not attempt re-identification.

Maintenance

Maintained by the JuDDGES organization at Wrocław University of Science and Technology. Versioned via Hugging Face commits; dataset card updates announce material changes.

Citation

@inproceedings{juddges-pl-2026,
  title     = {JuDDGES-pl: Hundreds of Thousands of Structurally Enriched Polish Judgments from Common and Administrative Courts},
  author    = {Augustyniak, {\L}ukasz and others},
  booktitle = {NeurIPS 2026 Evaluations and Datasets Track},
  year      = {2026}
}

Croissant metadata

Croissant JSON-LD with full Responsible AI fields is released alongside the paper (NeurIPS 2026 supplementary material).

Downloads last month
332