This dataset contains 428K parsed Ukrainian court decisions for temporal drift research. Access is gated to track usage. Please describe your intended use case.

UA-Temporal-Drift: Temporal Concept Drift Benchmark for Legal Judgment Prediction

A benchmark dataset of 428,075 Ukrainian court decisions spanning 2008–2026, designed to measure temporal concept drift in legal NLP models. Decisions are organized into three temporal epochs reflecting major geopolitical disruptions to the Ukrainian judicial system, with chronological train/validation/test splits within each epoch.

Motivation

Legal NLP benchmarks (LexGLUE, LEXTREME, SCALE) evaluate models on randomly split data, implicitly assuming legal language is stationary. This dataset enables cross-temporal evaluation: train on one epoch, test on another, to measure how model performance degrades over time.

Prior work established a 27.9 percentage-point forward degradation gap using TF-IDF classifiers on this data. This dataset supports neural experiments to determine whether transformer fine-tuning and legal-domain pretraining mitigate temporal drift.

Temporal Epochs

Epoch	Period	Decisions	Context
`pre_war`	2008–2013	128,075	Peacetime. All 832 courts operational, stable procedural rules.
`hybrid_war`	2014–2021	150,000	Crimea annexation (2014), judicial reform (2017), procedural modernization.
`full_scale`	2022–2026	150,000	Full-scale invasion, martial law, new Criminal Code articles, military case surge.

Task

Judgment prediction (3-class classification): given the facts section of a court decision, predict the outcome.

Label	Ukrainian	Meaning
`approved`	задоволено	Claim fully granted
`dismissed`	відмовлено	Claim denied
`partial`	частково задоволено	Claim partially granted

Dataset Structure

Configs

all (default) -- all 428K decisions combined, chronological split
pre_war -- 128,075 decisions from 2008–2013
hybrid_war -- 150,000 decisions from 2014–2021
full_scale -- 150,000 decisions from 2022–2026

Splits

Within each config, documents are split chronologically (not randomly):

train (80%) -- earliest decisions
validation (10%) -- middle
test (10%) -- most recent decisions

This prevents temporal leakage: models are always evaluated on decisions that postdate their training data.

Fields

Field	Type	Description
`text`	string	Facts section of the court decision (model input). PII replaced with `[PERSON]`, `[ADDRESS]`, `[NUMBER]`, `[INFO]`. Truncated to 10,000 characters.
`label`	string	Outcome: `approved`, `dismissed`, or `partial`
`epoch`	string	Temporal epoch: `pre_war`, `hybrid_war`, or `full_scale`
`adjudication_date`	string	Date of the decision (YYYY-MM-DD)
`jurisdiction`	string	`civil` or `commercial`
`doc_id`	string	EDRSR document ID (links to https://reyestr.court.gov.ua/)
`language`	string	Always `uk` (Ukrainian)

Statistics

Config	Split	Samples	Avg length (chars)
pre_war	train	102,460	5,267
pre_war	validation	12,807	4,773
pre_war	test	12,808	5,324
hybrid_war	train	120,000	6,456
hybrid_war	validation	15,000	7,698
hybrid_war	test	15,000	7,783
full_scale	train	120,000	8,068
full_scale	validation	15,000	8,678
full_scale	test	15,000	8,731
Total		428,075

Note: average text length increases across epochs (5.3K → 8.7K chars), reflecting increasing judicial documentation standards.

Usage

from datasets import load_dataset

# Load a single epoch
ds = load_dataset("overthelex/ua-temporal-drift", "hybrid_war")

# Cross-epoch evaluation: train on pre_war, test on full_scale
train = load_dataset("overthelex/ua-temporal-drift", "pre_war", split="train")
test = load_dataset("overthelex/ua-temporal-drift", "full_scale", split="test")

# Load all epochs combined
ds = load_dataset("overthelex/ua-temporal-drift", "all")

Cross-Temporal Evaluation Protocol

from datasets import load_dataset
from transformers import AutoModelForSequenceClassification, Trainer

epochs = ["pre_war", "hybrid_war", "full_scale"]

# Train one model per epoch
for train_epoch in epochs:
    train_ds = load_dataset("overthelex/ua-temporal-drift", train_epoch, split="train")
    model = AutoModelForSequenceClassification.from_pretrained("xlm-roberta-base", num_labels=3)
    # ... fine-tune on train_ds ...

    # Evaluate on ALL epochs
    for test_epoch in epochs:
        test_ds = load_dataset("overthelex/ua-temporal-drift", test_epoch, split="test")
        # ... evaluate model on test_ds ...
        # This produces a 3x3 generalization matrix

Source

Decisions are extracted from the Unified State Register of Court Decisions (EDRSR / ЄДРСР), a publicly accessible database of all Ukrainian court decisions since 2006 containing over 100 million documents.

Registry: https://reyestr.court.gov.ua/
Individual decisions can be accessed at: https://reyestr.court.gov.ua/Review/{doc_id}

Related Datasets

overthelex/ukrainian-court-decisions -- 7-class judgment prediction with 14,452 decisions
overthelex/ua-case-outcome -- case outcome prediction benchmark
overthelex/ua-legal-bench -- LLM evaluation on Ukrainian legal tasks
overthelex/ua-statute-retrieval -- statute retrieval benchmark
joelniklaus/lextreme -- multilingual legal NLP benchmark (submitted as Ukrainian contribution)

Citation

@misc{ovcharov2025temporaldrift,
  title={Temporal Concept Drift in Legal Judgment Prediction: Neural Baselines Across Three Epochs of Ukrainian Court Decisions},
  author={Ovcharov, Volodymyr},
  year={2025},
  url={https://huggingface.co/datasets/overthelex/ua-temporal-drift},
}

Access

This dataset is gated: you must request access and describe your use case. Requests are reviewed manually. Academic and research use is approved by default.

License

CC-BY-4.0. The source data (EDRSR) is publicly available under Ukrainian open data legislation.

Downloads last month: 20

Total file size:

3.98 GB

You need to agree to share your contact information to access this dataset