---
language:
  - pl

pretty_name: ECtHR-PL-AVA

task_categories:
  - text-classification

size_categories:
  - 1K<n<10K

tags:
  - polish
  - legal
  - long-context
  - document-classification
  - multi-label-classification
  - human-rights
  - ecthr
  - translated

dataset_info:
  features:
    - name: text
      dtype: string
    - name: labels
      sequence:
        dtype: int64
  splits:
    - name: train
      num_examples: 7367
    - name: validation
      num_examples: 834
    - name: test
      num_examples: 819
---
# ECtHR-PL-AVA

**ECtHR-PL-AVA** is a Polish long-document multi-label classification dataset
derived from the publicly available
[ECtHR Cases](https://huggingface.co/datasets/AUEB-NLP/ecthr_cases) dataset.

The dataset contains Polish translations of factual descriptions from cases
heard by the European Court of Human Rights (ECtHR). The task is to predict
the **articles of the European Convention on Human Rights alleged to have
been violated** in a given case.

ECtHR-PL-AVA is part of the **LongContext** benchmark introduced with
[Polish ModernBERT](ARXIV_URL).

## Dataset statistics

| Split | Examples |
|---|---:|
| Train | 7,367 |
| Validation | 834 |
| Test | 819 |
| **Total** | **9,020** |

The dataset contains predominantly long documents: **83.4%** of test examples
exceed 512 tokens, **58.6%** exceed 1,024 tokens, and **23.3%** exceed
2,048 tokens when tokenized with the Polish ModernBERT Base tokenizer.

## Dataset construction

For each case, sentences and paragraphs describing the case facts were
concatenated into a single document. Cases with factual descriptions longer
than 32,000 characters were removed to exclude extreme-length outliers.

The remaining documents were translated from English into Polish using
**GLM-4.6**, while the original dataset splits and labels were retained.

The task is evaluated using **weighted F1**.

## Citation

If you use this dataset, please cite:

[CITATION]