Abstract
Existing legal benchmarks including LegalBench (NeurIPS 2023) and Harvey's LAB (2026) are built exclusively on US and UK law. Neither covers Africa's 54 jurisdictions, pluralistic legal systems, or multilingual proceedings. LegalBench explicitly notes that it skews toward US Federal law and includes no multilingual or non-English tasks. This gap is not a minor omission; it means that legal AI systems evaluated solely on these benchmarks have never been tested against the legal traditions that govern hundreds of millions of people across the African continent. AfriLegalBench addresses this gap by providing a structured evaluation suite for legal reasoning in African legal contexts. Phase 1 covers three jurisdictions spanning distinct legal traditions. Nigeria, with 20 tasks and 62 examples, operates under a hybrid system of English common law, customary law, and Sharia law across its 36 states. South Africa, with 7 tasks and 20 examples, combines Roman-Dutch law, English common law, and customary law across 11 official languages. Kenya, with 6 tasks and 17 examples, blends English common law with customary and Islamic law in English and Swahili. The benchmark extends LegalBench's taxonomy with four additional reasoning categories specifically designed for African legal contexts: customary law reasoning, Sharia law reasoning, multilingual statute retrieval, and legal system identification. It also covers statute interpretation, case citation verification, long document understanding, multi-jurisdictional citation analysis, and analogical reasoning across jurisdictions. Each task is a JSON file conforming to the task schema and compatible with existing LegalBench evaluation pipelines, enabling direct comparison with US and UK results. Data sources include SAFLII, NigeriaLII, and KenyaLII, all open access case law repositories.