ECtHR-NPD — unified current release

ECtHR-NPD is a benchmark for predicting non-pecuniary damage awards at the European Court of Human Rights. It contains 14,575 cases with chronological train/validation/test partitions and ID, OOD and Challenging test views.

Data

• data/case_level.csv: 14,575 cases; 33 columns.
• data/applicant_level.csv: 44,581 applicant source units; 14 columns.
• Join on case_id, the original HUDOC item identifier. applicant_id remains a stable source-record key, not an independently identified natural person.
• Train / validation / test: 10,217 / 1,461 / 2,897. ID / OOD: 1,000 / 1,897. Challenging: 699, overlapping the test pool, not a fourth physical split.
• Names are [MASKED]. Dates, exact amounts, birth years, nationality text and other structured attributes remain. This is name-masked, not anonymous.
• No ratio, one-hot, log-transformed or numerically encoded predictor columns. Amounts, years, counts and durations are natural numeric values. y_binary is the existing outcome label, not a predictor encoding.
• unknown is missing or unresolved, never an invented zero. Estate awards are identified by npd_award_scope; no amount is divided among heirs or groups.

split supplies logical partitions in each CSV. Hugging Face config all means the full physical table, not the training split. See CSV_TREE.md for every column and DATA_DICTIONARY.md for meanings.

Reproduction entry points

Run from this folder using Python 3.10+; set PYTHONDONTWRITEBYTECODE=1 if desired.

python -B code/public_tables.py .
python -B -m unittest discover -s tests

Tests require the dependencies appropriate to their tested components. Install requirements-extraction.txt for extraction, and consult the experiment guide for optional modelling environments rather than assuming all GPU packages are needed to read the CSVs. The table validator itself uses the standard library.

• Extraction and source reconstruction: use the released HUDOC IDs to obtain sources; create extraction candidates; run explicit review checks before exporting two readable tables.
• Experiments: prepare model-only encodings locally, fit/select on train/validation, evaluate on fixed test views, and keep target evidence out of predictors. The public CSVs themselves remain unencoded.
• Paper alignment and limits: separates implemented methods, tested behaviours, missing historical evidence and unrun experiments.
• Validation record: tests and actual checks for this ZIP.

Original checkpoints, predictions, provider responses, exact historical feature maps and complete manual-review ledgers have not all been recovered. This code supports transparent new runs; it does not authenticate the paper's reported scores. New readable-feature projections must not be called the original X1.

Only two CSVs are distributed. Raw judgments, name mappings, review evidence and API logs are not bundled; local source reconstruction may create identifying files and send material to a configured provider only on explicit invocation. The extraction/experiment workflow does not upload its generated artifacts. Legacy internal-contract helpers retained for regression testing do not define extra public releases.

Maintenance and reuse

Quarterly updates are intended to add newly eligible cases after review, with a dated change log and content hashes. This is not a claim of an unattended service or that the March 2026 source snapshot has already been extended. Cite the paper and the exact content revision used in your research. Please report suspected extraction errors.

See LICENSE_AND_SOURCE_TERMS.md: this distribution does not select or grant a new open licence, nor assert compliance approval. No personal-name mapping is needed or included to join the two tables.

Where to review this release

• GitHub: https://github.com/YanyiPU/ECtHR-NPD
• Hugging Face: https://huggingface.co/datasets/YanyiPU716/ECtHR-NPD

The dataset tables and research source code are available from these repositories. Use the data and code together and record the revision used in your research.