Multi-Legal-Bench

Paper: Multi-Legal-Bench: Evaluating LLMs on Legal Reasoning Across Jurisdictions, Languages, and Legal Traditions

The first cross-jurisdictional legal benchmark that evaluates identical tasks across six countries, five language families, and 122 million court decisions.

Overview

Multi-Legal-Bench defines five tasks mapped to structured metadata from national court registries:

Task	Description	Metric
CTC	Court-Type Classification	Accuracy
JFC	Judgment Form Classification	Accuracy
COP	Case-Outcome Prediction	Accuracy
CCP	Cause Category Prediction	Accuracy
NE	Legal Norm Extraction	Set-level F1

Jurisdictions

Code	Country	Language	Family	Full-text decisions
UA	Ukraine	Ukrainian	Slavic	100.8M
FR	France	French	Romance	719K
NL	Netherlands	Dutch	Germanic	921K
PL	Poland	Polish	Slavic	2.8M
CZ	Czech Republic	Czech	Slavic	871K
LT	Lithuania	Lithuanian	Baltic	51K

Task-Jurisdiction Coverage

Not every task is available in every jurisdiction (metadata availability varies):

Task	UA	FR	NL	PL	CZ	LT
CTC	5/5	3	6	6	3	4
JFC	4	6	2	3	3	--
COP	6	7	--	--	--	--
CCP	17	10	11	--	--	10
NE	yes	--	--	yes	yes	--

Data Format

Each JSONL file contains records with:

• doc_id: unique decision identifier
• jurisdiction: country code (FR, NL, PL, CZ, LT)
• language: language code
• task: task identifier (ctc, jfc, cop, ccp, ne)
• label_original: original metadata label
• label_normalized: normalized label for evaluation
• decision_date: date of the decision
• text_length: character count
• full_text: complete decision text
• facts: extracted facts section (COP only)
• norms: extracted legal norm references (NE only)

Files

data/
  fr_ctc.jsonl   # 999 French court-type classification
  fr_jfc.jsonl   # 790 French judgment form classification
  fr_cop.jsonl   # 813 French case-outcome prediction
  fr_ccp.jsonl   # 1000 French cause category prediction
  nl_ctc.jsonl   # 996 Dutch court-type classification
  nl_jfc.jsonl   # 1000 Dutch judgment form classification
  nl_ccp.jsonl   # 1000 Dutch cause category prediction
  pl_ctc.jsonl   # 498 Polish court-type classification
  pl_jfc.jsonl   # 999 Polish judgment form classification
  pl_ne.jsonl    # 999 Polish norm extraction
  cz_ctc.jsonl   # 333 Czech court-type classification
  cz_jfc.jsonl   # 669 Czech judgment form classification
  cz_ne.jsonl    # 1000 Czech norm extraction
  lt_ctc.jsonl   # 1000 Lithuanian court-type classification
  lt_ccp.jsonl   # 1000 Lithuanian cause category prediction

Total: 13,096 decisions across 15 task-jurisdiction combinations.

Key Findings

Evaluated on 11 LLMs (7 frontier + 4 small) via AWS Bedrock:

1. Task difficulty is stable across jurisdictions: CTC near-ceiling (52--100%), COP genuinely hard (30--42%)
2. Few-shot effects are task-dependent, not language-dependent: JFC benefits everywhere (+3.7 to +66.2pp), COP mixed everywhere
3. Cross-lingual transfer depends on label alignment, not language proximity: UA->FR (-2.1pp) outperforms UA->PL (-13.7pp)
4. Tokenizer fertility does not predict cross-lingual accuracy: r=-0.27, p=0.14 (n.s.)
5. Small models (3-12B) degrade most on under-represented languages: LT CTC gap -20pp

Papers

This dataset is used in the following papers:

• Multi-Legal-Bench: Evaluating LLMs on Legal Reasoning Across Jurisdictions, Languages, and Legal Traditions -- arXiv:2605.29738 [cs.CL]
• UA-Legal-Bench: A Benchmark for Evaluating Large Language Models on Ukrainian Legal Reasoning -- arXiv:2605.29170 [cs.CL]

Citation

@article{ovcharov2026multilegalbench,
  title={Multi-Legal-Bench: Evaluating LLMs on Legal Reasoning Across Jurisdictions, Languages, and Legal Traditions},
  author={Ovcharov, Volodymyr},
  journal={arXiv preprint arXiv:2605.29738},
  year={2026},
  url={https://arxiv.org/abs/2605.29738}
}

@article{ovcharov2026ualegalbench,
  title={UA-Legal-Bench: A Benchmark for Evaluating Large Language Models on Ukrainian Legal Reasoning},
  author={Ovcharov, Volodymyr},
  journal={arXiv preprint arXiv:2605.29170},
  year={2026},
  url={https://arxiv.org/abs/2605.29170}
}

Data Sources

All data from official open court registries:

• FR: Cour de cassation Open Data (data.gouv.fr)
• NL: Rechtspraak.nl
• PL: Sadownictwo powszechne / NSA (orzeczenia.ms.gov.pl)
• CZ: Ustavni soud + Nejvy soud (nalus.usoud.cz)
• LT: LITEKO / e-teismai (liteko.teismai.lt)

License

CC-BY-SA 4.0