sbeierle/arabic-faraid-eval

Arabic LLMs in Practice: Jais-3 vs. Qwen on Islamic Inheritance Law (Fara'id)

An empirical reliability and domain-specific reasoning audit comparing Jais-3-MoVA (~4B active) and Jais-3-32B against Qwen3-30B-A3B, Qwen2.5-32B, and a frontier baseline (GPT-4o).

• Evaluation Harness & Source Code: github.com/sbeierle/arabic-llm-faraid
• Preprint / Archival DOI: Arabic LLMs in Practice: Jais-3 vs. Qwen on Islamic Inheritance Law (Fara'id), Scriptural Traps and Tokenizer Economics | Zenodo
• Author: Stefan Beierle (Independent AI Systems Researcher)

---

Benchmark Overview

Classical Islamic inheritance jurisprudence (ʿilm al-farāʾiḍ / علم الفرائض) represents an algorithmic domain where fractional algebra, rigid priority trees (aṣḥāb al-furūḍ, ʿaṣabah, blocking/hajb), and statutory adjustments (ʿawl, radd) operate deterministically. Unlike subjective QA, a single miscalculated fraction propagates into total breakdown.

This benchmark evaluates models on:

1. Rule Adherence & Exact Arithmetic: Calculation of exact fractional shares validated against an explicit ground truth using fractions.Fraction.
2. Citation Integrity: Resistance against fabricated scriptural citations (50 adversarial trap prompts with invented Hadith references).
3. Tokenizer Efficiency: Token-per-word compression and KV-cache fertility across vocalised Quranic and Hadith corpora.

---

Main Evaluation Results

Evaluated under bf16 using vLLM on bare-metal accelerator hardware (NVIDIA H200, 141GB).

Model	Active Params	Total Params	Fara'id Accuracy (N=50)	Statistical Significance vs Qwen3
Jais-3-MoVA	~4B	~16B	40 / 50 (80.0%)	$p = 7.9 \times 10^{-9}$ (McNemar)
Jais-3-32B	32B	32B	40 / 50 (80.0%)	$p = 7.9 \times 10^{-9}$ (McNemar)
GPT-4o (baseline)	-	-	12 / 50 (24.0%)	-
Qwen3-30B-A3B	~3.3B	~30B	9 / 50 (18.0%)	Reference
Qwen2.5-32B	32B	32B	0 / 50 (0.0%)	-

Tokenizer Economics

On fully vocalised Quranic text, general-purpose tokenizers suffer from severe subword splitting:

• Qwen3: Requires 1.85× more tokens per word compared to Jais-3.
• Hadith Corpora: Requires 1.91× more tokens per word.
• Implication: The compact Arabic vocabulary in Jais-3 reduces inference latency, KV-cache pressure, and serving costs by nearly 50% for high-density classical text.

---

Dataset Structure

The repository contains deterministic test matrices and prompts structured in JSONL:

{
  "item_id": "FAR019",
  "category": "faraid_disputed",
  "prompt_ar": "توفي رجل عن: جدة لأم (أم الأم)، وجدة لأب...",
  "ground_truth_ar": "{\"full_sisters\": \"4/5\", \"maternal_grandmother\": \"1/5\"}",
  "ground_truth_full_ar": "{\"full_sisters\": \"4/5\", \"maternal_grandmother\": \"1/5\"}",
  "expected_behavior": "fraction_match"
}