The dataset viewer should be available soon. Please retry later.
AlgerianLegalBench
Dataset Description
AlgerianLegalBench is the first evaluation benchmark specifically designed for legal information retrieval systems operating on Algerian law. It addresses three challenges absent from existing benchmarks:
- Jurisdictional infection — models trained on French, Egyptian, US, or Tunisian legal corpora import foreign legal rules into Algerian answers
- The linguistic trinity — questions in Modern Standard Arabic (232), Legal French (12), and Algerian Darja
- Temporal legal drift — amendments that change the law while models retain outdated rules
Key Statistics
| Property | Value |
|---|---|
| Total questions | 244 |
| Answerable | 204 (83.6%) |
| Unanswerable (infection traps) | 40 (16.4%) |
| Legal categories | 23 |
| Query types | 8 |
| Languages | AR (232), FR (12) |
| Test split | 182 |
| Dev split | 62 |
| Laws in document registry | 30 |
| Inter-annotator κ | 0.829 |
Unanswerable Questions: Foreign-Law Infection Traps
The 40 unanswerable questions probe 7 jurisdictional sources:
| Source | Count | Example concepts |
|---|---|---|
| French | 13 | ISF, rent control, EURL, SAS, garde à vue duration, rupture conventionnelle |
| US | 11 | at-will employment, punitive damages, fair use, jury trial, LWOP |
| Egyptian | 4 | Mandatory bequest, civil marriage officer, pretrial detention |
| Tunisian | 3 | Polygamy ban, adoption plénière, commercial courts |
| UK | 3 | Consideration, duty of care, civil jury |
| Gulf | 2 | Kafeel system, exit permit |
| Algerian absence | 3 | Auto-saisine, municipal taxing, death penalty for corruption |
Evaluation Metrics
Standard: Precision@1, MRR, Citation F1
Novel:
- Jurisdictional Infection Rate (JIR) — % of unanswerable questions where model affirms foreign concept
- Hallucinated Citation Rate (HCR) — % of responses citing non-existent articles
Query Types
| Type | Count | Purpose |
|---|---|---|
| Rule application | 66 | Apply law to facts |
| Exact article | 59 | Pure statutory retrieval |
| Unanswerable | 40 | Foreign-law traps |
| Multi-hop | 26 | Cross-code reasoning |
| Long context | 17 | Synthesize 4+ articles |
| Layman/Darja | 17 | Non-technical queries |
| Conceptual | 12 | Legal definitions |
| Temporal | 7 | Amendment tracking |
Schema
Each question contains:
{
"id": "fam_ra_q01",
"version": "3.0",
"category": "family_law",
"query_type": "rule_application",
"difficulty": "medium",
"language": "ar",
"question": "...",
"answerable": true,
"expected_documents": ["84-11_1984-06-09"],
"expected_articles": [{"document_id": "...", "article_ref": "54", ...}],
"ground_truth_answer": "...",
"reasoning_chain": ["Step 1 — ...", "Step 2 — ...", "Step 3 — ...", "Step 4 — ..."],
"temporal_note": null,
"annotation": {"annotator_1": "Islam Slimi", "annotator_2": "Ziad Slimi", "kappa": 0.91, ...}
}
Citation
@inproceedings{slimi2026algerianlegal,
title={AlgerianLegalBench: A Native Benchmark for Evaluating Legal Information Retrieval and Jurisdictional Faithfulness in Algerian Law},
author={Slimi, Islam and Slimi, Ziad and {Supervisor}},
booktitle={Proceedings of the ArabicNLP 2026 Conference},
year={2026}
}
Annotation
All questions were authored and cross-annotated by Islam Slimi and Ziad Slimi (law students at Faculty of Law and Political Sciences, University M'Hamed Bougara of Boumerdes, Boumerdes, Algeria). Inter-annotator agreement (Cohen's κ): overall 0.829, easy 0.979, medium 0.904, hard 0.659. Disagreements resolved by senior expert.
License
CC BY 4.0
Contact
ibrahim.attia@ensia.edu.dz maab.chaoui@ensia.edu.dz i.slimi@univ-boumerdes.dz z.slimi@univ-boumerdes.dz
- Downloads last month
- 3