QU-NLP at QIAS 2026: Multi-Stage QLoRA Fine-Tuning for Arabic Islamic Inheritance Reasoning

Mohammad AL-Smadi

cs.CLMar 29, 2026

QU-NLP at QIAS 2026: Multi-Stage QLoRA Fine-Tuning for Arabic Islamic Inheritance Reasoning

Authors: Mohammad AL-Smadi

Organizations: Qatar University Doha, Qatar

Abstract

Islamic inheritance law (ilm al-mawarıth) presents a challenging domain for evaluating large language models' structured reasoning capabilities, requiring multi-step legal analysis, rule-based blocking decisions, and precise fractional calculations. We present QU-NLP's submission to the QIAS 2026 shared task on Arabic Islamic inheritance reasoning. Our approach employs a multi-stage Quantized Low-Rank Adaptation (QLoRA) fine-tuning strategy on Qwen3-4B: (1) domain adaptation on 3,166 Islamic fatwa records to acquire inheritance terminology and jurisprudential reasoning patterns, followed by (2) task-specific training on 12,000 structured inheritance cases to optimize JSON-formatted output generation. Using 4-bit NF4 quantization with rank-128 LoRA adapters, our model achieves 90% MIR-E (Mawarith Inheritance Reasoning Evaluation) score on the test set, demonstrating competitive performance while requiring minimal computational resources. Our results show that domain-specific pre-adaptation combined with structured output training enables small language models to perform complex legal reasoning tasks effectively comparing to commercial systems such as Gemini-2.5-flash.

Figures & tables

Split	Simple	’awl	radd	Total
Training	11,079	577	344	12,000
Test	456	39	5	500
Total	11,535	616	349	12,500

Table 1: MAWARITH dataset structure with distribution of inheritance cases by complexity.

Field	Content
Question	\textAR مات وترك: عم لأب و ابن أخ لأب و أربع بنات ابن و أم الأم و أب الأب و زوجة و خمسة أبناء ابن أخ لأب. ما هو نصيب كل وريث؟
	(A person died leaving: a paternal uncle, a son of a paternal brother, four sons’ daughters, the maternal grandmother, the paternal grandfather, a wife, and five sons of a paternal nephew. What is each heir’s share?)
Heirs	[ \textAR وريث: أم الأم، عدد: 1، \textAR وريث: زوجة، عدد: 1، \textAR وريث: أب الأب، عدد: 1، \textAR وريث: بنت ابن، عدد: 4 ]
	(heir: maternal grandmother, count: 1; heir: wife, count: 1; heir: paternal grandfather, count: 1; heir: son’s daughter, count: 4)
Blocked	[ \textAR وريث: عم لأب، عدد: 1، \textAR وريث: ابن أخ لأب، عدد: 1، \textAR وريث: ابن ابن أخ لأب، عدد: 5 ]
	(heir: paternal uncle, count: 1; heir: son of a paternal brother, count: 1; heir: sons of a son of a paternal brother, count: 5)

Table 2: Example inheritance case from MAWARITH with Arabic text and full English translation. The prescribed shares exceed the estate, so the case undergoes \textAR عول (proportional reduction). The four \textAR بنات ابن (son’s daughter) share 16/27 collectively, giving each one 4/27.

Configuration	Phase 1: Domain Adaptation	Phase 2: Task-Specific Training
Training data	3,166 Islamic fatwa records	12,000 inheritance cases
Objective	Causal language modeling	Structured JSON generation
Learning rate	$2\times 10^{-4}$	$3\times 10^{-5}$
Epochs	2	6
Batch size	1	1
Gradient accumulation	8 steps	8 steps

Table 3: Training configuration for the two phases of the proposed multi-stage QLoRA fine-tuning approach. Phase 1 performs domain adaptation on Islamic fatwa records, while Phase 2 focuses on structured inheritance reasoning and JSON output generation.

Model	$S_{h}$	$S_{s}$	$S_{a}$	MIR-E
Gemini-2.5-flash	94.5%	92.9%	89.4%	90.1%
QU-NLP (Ours)	97.1%	94.3%	84.6%	90.0%
Qwen3-32B	69.0%	44.6%	26.5%	43.7%
GPT-OSS-120B	69.3%	32.7%	27.1%	39.1%
LLaMA-3.3-70B	64.8%	40.3%	21.5%	39.0%
Fanar-Sadiq	62.1%	36.7%	20.4%	36.8%

Table 4: Component-wise MIR-E scores on the 500-case test set after post-processing.

S_{h}

: heirs and blocking,

S_{s}

: share assignment,

S_{a}

: adjustment detection. QU-NLP results reflect the Basic post-processing pipeline.

Metric	Original	Basic	PostTasil
MIR-E Overall	89.8%	90.0%	90.0%
Heirs & Blocking	96.4%	97.1%	97.1%
Share Assignment	94.3%	94.3%	94.3%
’Awl/Radd Detection	84.6%	84.6%	84.6%
Final Distribution	80.5%	80.5%	80.5%

Table 5: Effect of post-processing on component scores across 500 test cases. Original : raw model output. Basic : typographic cleaning and structural deduplication. PostTasil : Basic plus recalculation of final distributions for \textAR عول/ \textAR رد cases. Bold denotes the highest score per metric.

Type	N	MIR-E	$S_{h}$	$S_{s}$	$S_{a}$	$S_{f}$
Simple	456	90.8%	96.4%	94.1%	86.2%	83.5%
\textAR عول ( ’awl )	39	79.2%	96.2%	95.1%	66.7%	50.4%
\textAR رد ( radd )	5	83.0%	100.0%	100.0%	80.0%	50.0%

Table 6: Per-category MIR-E scores. ’Awl cases underperform radd cases (79.2% vs. 83.0%) despite nearly eight times more training examples, revealing arithmetic complexity rather than data frequency as the primary bottleneck.

Error Type	Cases	Total	Impact
Calculation	29	5.8%	$-$ 1.7pp
Residue label avoidance †	314	62.9%	$-$ 0.85pp
Heir Identification	19	3.8%	$-$ 0.4pp
Radd Detection	11	2.2%	$-$ 0.2pp

Table 7: Distribution of error types with impact on overall MIR-E. Calculation, heir identification, and radd detection are discrete failure cases (59 total, 11.8% of test set); impacts are estimated. † Among the 417 cases where the gold solution requires \textAR باقى التركة, the model substitutes an explicit fraction in 314 (75.3%). Because 83.1% of those 314 cases write the numerically correct fraction, MIR-E’s tolerance absorbs most of the penalty: the measured shares score gap between cases with and without the label is 0.045, giving a global cost of 314/500 x 0.045 x 0.30 =

-0.85

pp (shares score: 0.97 with label vs. 0.92 without, gap = 0.045). This row is not summed with the three discrete error categories above.

Case ID	Error Type	$S_{h}$	$S_{s}$	$S_{a}$	$S_{f}$	MIR-E
nf8w4p3x_7	Calculation	1.00	1.00	1.00	0.00	0.700
nf5n5k1z_7	Heir ID	0.66	1.00	1.00	1.00	0.898
ng7z4j2b_2	Radd Detection	1.00	1.00	0.00	0.00	0.600
ng2p5t2e_10	Share Assignment	0.81	0.67	0.00	0.33	0.542

Table 8: Component scores for representative error cases (

S_{h}

: heirs/blocking,

S_{s}

: shares,

S_{a}

: adjustment,

S_{f}

: final distribution).

Residue label behaviour	Cases	Rate
Gold requires residue label	417	83.4% of test
Model provides label (recall)	103	24.7% of required
Model avoids label	314	75.3% of required
of which: correct fraction	261	83.1% of avoided
of which: wrong fraction	53	16.9% of avoided
Global MIR-E cost	$-0.85$ pp

Table 9: Residue label recall analysis. Despite a 75.3% avoidance rate, the global MIR-E cost is only

-0.85

pp because 83.1% of avoidance cases compute the numerically correct fraction within the evaluation tolerance.

Reasoning Stage	Success Rate
$S_{h}=1$ (Heirs correct)	84.0%
$S_{h}=1,\ S_{s}=1$ (+ Shares correct)	81.2%
$S_{h}=1,\ S_{s}=1,\ S_{a}=1$ (+ Adjustment)	79.4%
All stages correct	65.5%

Table 10: Cumulative pipeline success rates. Each row shows the percentage of cases where all stages up to and including that point score perfectly (

S=1

Legal Reasoning

Legal reasoning studies how people and AI interpret laws, apply rules to facts, weigh evidence, and justify legal conclusions.

See all 104 papers

QIAS 2026: Overview of the Shared Task on Islamic Inheritance Reasoning
Jun 11, 2026Abdessalam Bouchekif, Somaya Eltanbouly, Samer Rashwani +5LLM Evaluation Legal Reasoning
Hamad bin Khalifa University, Qatar · Nazarbayev University, Kazakhstan
Which Models Perform Better in Inheritance Reasoning?
Jun 11, 2026Mohammed Amine Mouhoub, Chahinez BouchekifLLM Evaluation LLM Reliability
Paris Dauphine University · University of Abou Bekr Belkaïd
MAWARITH: A Dataset and Benchmark for Legal Inheritance Reasoning with LLMs
Mar 8, 2026Abdessalam Bouchekif, Shahd Gaben, Samer Rashwani +5Legal Reasoning LLM Evaluation
Hamad Bin Khalifa University, Qatar · Nazarbayev University, Kazakhstan

Jun 11, 2026J/K move · Enter open · S save

QIAS 2026: Overview of the Shared Task on Islamic Inheritance Reasoning

Abdessalam Bouchekif, Somaya Eltanbouly, Samer Rashwani, Shahd Gaben+4

Hamad bin Khalifa University, Qatar · Nazarbayev University, Kazakhstan

LLM Evaluation Legal Reasoning

Open paper AskPDF arXiv

This paper presents a comprehensive overview of the QIAS 2026 shared task, organized as part of the OSACT7 Workshop and co-located with LREC 2026. The shared task was designed to evaluate the ability of large language models to perform complex reasoning in the religious and legal domain of Islamic inheritance. Unlike conventional question-answering benchmarks, QIAS 2026 focuses on end-to-end reasoning from natural language cases, requiring systems to perform the full inheritance calculation process, from identifying the eligible heirs to assigning the correct share to each beneficiary. To support this evaluation, the task was based on the MAWARITH benchmark, a dataset of

12{,}500

Arabic inheritance cases annotated with intermediate reasoning steps and final answers. System submissions were evaluated using MIR-E, a multi-step metric that measures performance across the main stages of inheritance reasoning. A total of

16

teams participated in the shared task, investigating a range of approaches, including prompting-based methods, retrieval-augmented generation, and fine-tuning strategies. The results show that Islamic inheritance remains a highly challenging benchmark for current language models, especially in stages that require precise legal interpretation and structured numerical reasoning. This overview summarizes the task design, dataset, evaluation framework, participating systems, and main results.

Abstract

Legal Reasoning

Explore similar work

QIAS 2026: Overview of the Shared Task on Islamic Inheritance Reasoning

Which Models Perform Better in Inheritance Reasoning?

MAWARITH: A Dataset and Benchmark for Legal Inheritance Reasoning with LLMs