Open Empirical Preprints & Reproducible Artifacts

Arabic Sovereign LLM & Jurisprudence Benchmark

Empirical evaluation of Arabic-centred foundation models (Jais-3, Falcon3) against multilingual architectures (Qwen, GPT-4o). Measuring deterministic Islamic inheritance law (علم الفرائض), scriptural completion, tokenizer economics on H200, and runtime hardware containment under UAE statutory law on NVIDIA Blackwell.

Zenodo Preprint v1.2 (Nov 2026) 1x NVIDIA H200 SXM
Arabic LLMs in Practice: Jais-3 vs. Qwen on Fara'id, Scriptural Traps and Tokenizer Economics
Zenodo Preprint v2.0 (Oct 2026) 1x Blackwell MIG (2g.48gb)
Falcon3-7B-Instruct under UAE Statutory Frameworks: Measurements & Rule-Based Containment

Inheritance Law (Fara'id) Benchmark Matrix (N=50 Tasks, NVIDIA H200 SXM, vLLM BF16)

Scoring: Exact fractions in Python (No LLM-as-judge)
Model Architecture Parameters / Class Fara'id Accuracy (N=50) Excl. 'Umarī Cases (N=44) Verse Completion (Tashkeel) Citation Traps (Clean / 50)
inception42/Jais-3-32B Dense (32B class) 40 / 50 (80.0%) 40 / 44 (90.9%) — —
inception42/Jais-3-MoVA-36B-A4B MoE (36B total / 4B active) 40 / 50 (80.0%) 40 / 44 (90.9%) 31 / 40 (77.5%) 26 / 50 (52.0%)
Qwen/Qwen3-30B-A3B MoE (30.5B total / 3.3B active) 9 / 50 (18.0%) 9 / 44 (20.5%) 0 / 40 (0.0%) 29 / 50 (58.0%)
GPT-4o (Commercial API Baseline) Closed API reference 12 / 50 (24.0%) — — —
Qwen/Qwen2.5-32B-Instruct Dense (32B class) 0 / 50 (0.0%) 0 / 44 (0.0%) — —
Exact McNemar test across MoE pair (Jais-3-MoVA vs. Qwen3-30B-A3B): 32 tasks solved exclusively by Jais, 1 exclusively by Qwen3 (p = 7.9 × 10⁻⁹).