Skip to content

Polish Judgment Dataset Curation

This script (curate_polish_dataset.py) implements the dataset curation pipeline for GitHub Issue #12 - curating ~6K Polish court judgments with topic coverage matching the UK corpus.

Prerequisites

Before running this script, ensure you have completed Issues #10 and #11:

  1. Issue #10: UK topic taxonomy must exist at data/uk_topics_taxonomy.json
  2. Issue #11: Cross-jurisdictional queries must exist at data/cross_jurisdictional_queries.json

Setup

  1. Install dependencies: 

    pip install -r scripts/requirements.txt

  2. Set your HuggingFace token: 

    export HF_TOKEN=your_huggingface_token_here

Usage

Basic Usage

# Curate 6,000 Polish judgments (default)
python scripts/curate_polish_dataset.py

Testing

# Test with smaller sample (faster, good for development)
python scripts/curate_polish_dataset.py --sample 100 --target 300

Custom Target

# Curate different number of documents
python scripts/curate_polish_dataset.py --target 3000

Outputs

The script creates two main files in the data/ directory:

1. polish_judgments_6k.parquet

Curated dataset with columns: - id: Unique document identifier - text: Full judgment text - court: Court name - date: Judgment date - topic_primary: Primary topic label (e.g., "Criminal Sentencing") - topic_secondary: Legal category (e.g., "Criminal Law") - topic_id: Numeric topic ID matching UK taxonomy - source_dataset: Source HuggingFace dataset - query_used: Polish queries used for retrieval - relevance_score: Document relevance score (0.0-1.0) - signature: Case signature/sygnatura for deduplication

2. polish_dataset_stats.json

Curation statistics including: - Topic distribution (UK vs Polish counts) - Source dataset breakdown - Coverage gaps and shortfalls - Per-topic statistics

Data Sources

The script attempts to load from multiple HuggingFace datasets in priority order:

  1. Primary: JuDDGES/pl-court-raw-enriched (has factual_state and legal_state fields)
  2. Secondary: JuDDGES/pl-nsa-enriched (administrative courts)
  3. Fallback: JuDDGES/pl-court-raw, JuDDGES/pl-nsa

Algorithm

  1. Query-driven retrieval: Use Polish legal queries from Issue #11 to score documents for topic relevance
  2. Balanced sampling: Allocate documents proportionally to match UK topic distribution
  3. Deduplication: Remove duplicates based on case signatures
  4. Quality filtering: Filter out very short judgments (< 500 characters)

Performance Notes

  • Uses streaming mode to handle large datasets without memory issues
  • Processes documents in batches for efficient memory usage
  • Text preprocessing includes normalization for better Polish text matching

Troubleshooting

Missing Prerequisites

Error: UK topic taxonomy not found at data/uk_topics_taxonomy.json
Run the Issue #10 script first to generate the UK topic taxonomy.

Missing HF Token

Error: HF_TOKEN environment variable required
Set your HuggingFace token in the environment.

Low Coverage for Specific Topics

Check polish_dataset_stats.json for gaps. Some specialized topics may have limited Polish coverage in the available datasets.

Example Output

🏛️ Polish Judgment Curation
┌─ Polish Dataset Curation Summary ─┐
│                                   │
│ 📊 Dataset Overview               │
│    • Total curated documents: 5,847│
│    • Target document count: 6,000  │
│    • Topics with coverage: 5/5     │
│    • Minimum per topic: 50         │
│                                   │
│ 📈 Source Distribution            │
│    • pl-court-raw-enriched: 3,245 │
│    • pl-nsa-enriched: 2,602       │
│                                   │
│ 🎯 Coverage Quality               │
│    • Avg coverage ratio: 0.97      │
│    • Topics with gaps: 1           │
└───────────────────────────────────┘

This indicates successful curation with good coverage across all topics.