datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
ECHRThis is the ECHR dataset, a collection of 11.5K court cases extracted from the public database
of the European Court of Human Rights and further annotated by human experts. The dataset was
published along with this paper (pleae cite it
accordingly!) and can be donwloaded in its original form from this website.
Each instance in this dataset is a court case. Each court case is annotated with the following properties (the columns of the dataframe):
partition: a label indicating dataset… See the full description on the dataset page: https://huggingface.co/datasets/glnmario/ECHR.echrThe ECHR Cases dataset is designed for experimentation of neural judgment prediction, as in the original 2019 ACL paper "Neural Legal Judgment Prediction in English".echr-livehrb-temporal-1k
echr-livehrb-temporal-1k
Temporally distributed, outcome-balanced evaluation set for LiveHumanRightsBench.
1,212 ECtHR case–article instances drawn from 976 judgments, verdict removed.
Why this set exists
The earlier releases (echr-livehrb-static-2k, echr-livehrb-temporal-2k) preserve
the Court's natural base rate, which is 83.7% violation. That leaves only 327
no-violation cases in 2,000 — and measurement shows the interesting behaviour lives
almost entirely in… See the full description on the dataset page: https://huggingface.co/datasets/overthelex/echr-livehrb-temporal-1k.echr-livehrb-static-2k
echr-livehrb-static-2k
Balanced static evaluation set for LiveHumanRightsBench (ECtHR human-rights judgment prediction).
regular: 1000 pairs from overthelex/echr-verdict-free (Ukraine excluded), round-robin stratified by respondent country.
ukr: 1000 pairs from overthelex/echr-ukr-verdict-free, round-robin stratified by ECHR article.
Natural violation/no-violation distribution (base-rate balancing is a separate task). Seed=42, reproducible.
Verdict-free text… See the full description on the dataset page: https://huggingface.co/datasets/overthelex/echr-livehrb-static-2k.fr-echr
French European Court of Human Rights Dataset
Description
The European Court of Human Rights (ECtHR) adjudicates claims concerning infringements on human rights provisions outlined in the European Convention on Human Rights (ECHR) by European states. The Convention can be accessed at https://www.echr.coe.int/Documents/Convention_ENG.pdf. The dataset construction followed the methodology of Chalkidis et al. (2019), but focused on decisions available in French. This dataset… See the full description on the dataset page: https://huggingface.co/datasets/audibeal/fr-echr.echr-verdict-free
Changelog
v1.2 (2026-07-14) — added article_full:
New non-breaking column article_full with protocol-aware ECHR codes (e.g. P1-1, P4-2, P7-4), recovered by re-parsing the HUDOC conclusion. The legacy article field collapsed protocol prefixes (article == "1" is Article 1 of Protocol 1 / property, not Convention Article 1; article == "4" conflates Convention Art 4 and Protocol 4). article is unchanged for continuity; rows that cannot be resolved fall back to it. 2524 rows carry a… See the full description on the dataset page: https://huggingface.co/datasets/overthelex/echr-verdict-free.echr-livehrb-temporal-2k
echr-livehrb-temporal-2k
Temporally binned evaluation split for LiveHumanRightsBench (ECtHR
human-rights judgment prediction). Two temporal contamination-control axes,
built from verdict-free (contamination-controlled) ECtHR text.
regular_temporal (1000): ex-Ukraine cases from overthelex/echr-verdict-free,
binned by decision year over 2017-2026 (100/bin), round-robin stratified by
respondent country. Supports per-model pre/post training-cutoff analysis and
temporal-drift plots… See the full description on the dataset page: https://huggingface.co/datasets/overthelex/echr-livehrb-temporal-2k.echr-ukr-verdict-free
ECtHR Ukraine Verdict-Free Dataset
2,619 verdict-free case-article pairs from the European Court of Human Rights (ECtHR) involving Ukraine as respondent state. Each record contains the factual background, procedural history, parties' arguments, and cited legal standards -- with the Court's legal reasoning, findings, conclusions, and operative provisions surgically removed.
Dataset Description
This dataset is designed for evaluating LLM reliability in legal… See the full description on the dataset page: https://huggingface.co/datasets/overthelex/echr-ukr-verdict-free.ECHRChunkRetrieval