QIMMA قِمّة ⛰: A Quality-First Arabic LLM Leaderboard

Source: https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard Back to Articles

• 🔍 The Problem: Arabic NLP Evaluation Is Fragmented and Unvalidated
• ⛰ What’s in QIMMA?
• 🔬 The Quality Validation Pipeline- Stage 1: Multi-Model Automated Assessment - Stage 2: Human Annotation and Review
• ⚠️ What We Found: Systematic Quality Problems- By the Numbers - Taxonomy of Issues Found
• 💻 Code Benchmark: A Different Kind of Quality Work
• ⚙️ Evaluation Setup- Evaluation Framework - Metrics by Task Type - Prompt Templates
• 🏆 Leaderboard Results- The Size-Performance Relationship
• 🌟 What Makes QIMMA Different
• 🔗 Resources
• 🔖 Citation

QIMMA validates benchmarks before evaluating models, ensuring reported scores reflect genuine Arabic language capability in LLMs.

If you’ve been tracking Arabic LLM evaluation, you’ve probably noticed a growing tension: the number of benchmarks and leaderboards is expanding rapidly, butare we actually measuring what we think we’re measuring?

We builtQIMMAقمّة (Arabic for “summit”), to answer that question systematically. Instead of aggregating existing Arabic benchmarks as-is and running models on them, we applied a rigorous quality validation pipelinebeforeany evaluation took place. What we found was sobering: even widely-used, well-regarded Arabic benchmarks contain systematic quality issues that can quietly corrupt evaluation results.

This post walks through what QIMMA is, how we built it, what problems we found, and what the model rankings look like once you clean things up.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#%F0%9F%94%8D-the-problem-arabic-nlp-evaluation-is-fragmented-and-unvalidated🔍 The Problem: Arabic NLP Evaluation Is Fragmented and Unvalidated

Arabic is spoken by over 400 million people across diverse dialects and cultural contexts, yet the Arabic NLP evaluation landscape remains fragmented. A few key pain points have motivated this work:

**Translation issues.**Many Arabic benchmarks are translations from English. This introduces distributional shifts. Questions that feel natural in English become awkward or culturally misaligned in Arabic, making benchmark data less representative of how Arabic is naturally used.

**Absent quality validation.**EvennativeArabic benchmarks are often released without rigorous quality checks. Annotation inconsistencies, incorrect gold answers, encoding errors, and cultural bias in ground-truth labels have all been documented in established resources.

**Reproducibility gaps.**Evaluation scripts and per-sample outputs are rarely released publicly, making it hard to audit results or build on prior work.

**Coverage fragmentation.**Existing leaderboards cover isolated tasks and narrow domains, making holistic model assessment difficult.

To illustrate where QIMMA sits relative to existing platforms:

LeaderboardOpen SourceNative ArabicQuality ValidationCoding EvalPublic OutputsOALL v1✅Mixed❌❌✅OALL v2✅Mostly❌❌✅BALSAMPartial50%❌❌❌AraGen✅100%✅❌❌SILMA ABL✅100%✅❌✅ILMAAMPartial100%✅❌❌HELM Arabic✅Mixed❌❌✅⛰ QIMMA✅99%✅✅✅ QIMMA is the only platform combining all five properties: open source, predominantly native Arabic content, systematic quality validation, code evaluation, and public per-sample inference outputs.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#%E2%9B%B0-whats-in-qimma⛰ What’s in QIMMA?

QIMMA consolidates109 subsetsfrom14 source benchmarksinto a unified evaluation suite of over52,000 samples, spanning 7 domains:

DomainBenchmarksTask TypesCulturalAraDiCE-Culture, ArabCulture, PalmXMCQSTEMArabicMMLU, GAT, 3LM STEMMCQLegalArabLegalQA, MizanQAMCQ, QAMedicalMedArabiQ, MedAraBenchMCQ, QASafetyAraTrustMCQPoetry & LiteratureFannOrFlopQACoding3LM HumanEval+, 3LM MBPP+Code A few things stand out about this design:

• **99% native Arabic content.**The only exception is code evaluation, which is inherently language-agnostic.
• **First Arabic leaderboard with code evaluation.**QIMMA integrates Arabic-adapted versions of HumanEval+ and MBPP+, making it possible to assess coding capability with Arabic-language problem statements.
• **Diversity in Domains and Tasks.**QIMMA evaluates real-world competency areas including education, governance, healthcare, creative expression, and software development.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#%F0%9F%94%AC-the-quality-validation-pipeline🔬 The Quality Validation Pipeline

This is the methodological heart of QIMMA. Before running a single model, we applied amulti-stage validation pipelineto every sample in every benchmark.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#stage-1-multi-model-automated-assessmentStage 1: Multi-Model Automated Assessment

Each sample was independently evaluated by two state-of-the-art LLMs:

• Qwen3-235B-A22B-Instruct
• DeepSeek-V3-671B

We chose two models with strong Arabic capability but different training data compositions, so that theircombinedjudgment is more robust than either alone.

Each model scores a sample against a10-point rubric, with binary scores (0 or 1) per criterion:

A sample is eliminated if either model scores it below 7/10. Samples where both models agree on elimination are dropped immediately. However, where only one model flags a sample, it proceeds to human review in Stage 2.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#stage-2-human-annotation-and-reviewStage 2: Human Annotation and Review

Flagged samples are reviewed bynative Arabic speakerswith cultural and dialectal familiarity. Human annotators make final calls on:

• Cultural context and regional variation
• Dialectal nuance
• Subjective interpretation
• Subtle quality issues automated assessment may miss

For culturally sensitive content, multiple perspectives are considered, since “correctness” can genuinely vary across Arab regions.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#%E2%9A%A0%EF%B8%8F-what-we-found-systematic-quality-problems⚠️ What We Found: Systematic Quality Problems

The pipeline revealed recurring quality issues across benchmarks; not isolated errors, butsystematic patternsreflecting gaps in how benchmarks were originally constructed.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#by-the-numbersBy the Numbers

BenchmarkTotal SamplesDiscardedDiscard RateArabicMMLU14,1634363.1%MizanQA1,769412.3%PalmX3,001250.8%MedAraBench4,960330.7%FannOrFlop6,984430.6%ArabCulture3,48270.2%MedArabiQ49910.2%GAT13,9861~0.0%3LM STEM2,6091~0.0%AraDiCE-Culture18000.0%ArabLegalQA7900.0%AraTrust52200.0%

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#taxonomy-of-issues-foundTaxonomy of Issues Found

⚖️ Answer Quality

False or mismatched gold indices, factually wrong answers, missing or raw text answers.

📄 Text & Formatting Quality

Corrupt or illegible text, spelling and grammar errors, and duplicate samples.

💬 Cultural Sensitivity

Stereotype reinforcement and monolithic generalizations about diverse communities.

🤝 Gold Answer Compliance

Misalignment of gold answers with evaluation protocols.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#%F0%9F%92%BB-code-benchmark-a-different-kind-of-quality-work💻 Code Benchmark: A Different Kind of Quality Work

Code benchmarks required a different intervention. Rather than discarding samples, werefined the Arabic problem statementsin 3LM’s Arabic adaptations of HumanEval+ and MBPP+, leaving task identifiers, reference solutions, and test suites completely unchanged.

The modification rates were striking:

BenchmarkTotal PromptsModifiedUnchangedModification Rate3LM HumanEval+1641451988%3LM MBPP+3783087081% Modifications fell into five categories:

1. Linguistic refinement: normalizing toward natural Modern Standard Arabic and consistent imperative style
2. Clarity improvements: fixing ambiguous instructions and unclear constraints
3. Consistency normalization: standardizing mathematical terminology, punctuation, and example formatting
4. Structural corrections: fixing broken triple-quoted strings, indentation errors, corrupted text fragments
5. Semantic refinements: clarifying whether ranges are inclusive/exclusive, preserving task intent

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#%E2%9A%99%EF%B8%8F-evaluation-setup⚙️ Evaluation Setup

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#evaluation-frameworkEvaluation Framework

QIMMA usesLightEval,EvalPlusandFannOrFlopas its evaluation framework, chosen for consistency, multilingual community adoption, and reproducibility.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#metrics-by-task-typeMetrics by Task Type

Task TypeMetricBenchmarksMCQNormalized Log-Likelihood AccuracyAraDiCE-Culture, ArabicMMLU, ArabCulture, PalmX, 3LM STEM, MedArabiQ, GAT, MedAraBench, AraTrustMulti-select MCQProbability Mass on Gold ChoicesMizanQAGenerative QAF1 BERTScore (AraBERT v02)MedArabiQ, ArabLegalQA, FannOrFlopCodePass@13LM HumanEval+, 3LM MBPP+

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#prompt-templatesPrompt Templates

QIMMA standardizes prompting by question format, with six template types:

MCQ: generic multiple choice ·MCQ-C: multiple choice with context passage ·MCQ-I: multiple choice with specific instructions (GAT analogy/completion) ·QA: generic open-ended QA ·QA-C: QA with context ·QA-F: fill-in-the-blank QA All prompts are in Arabic. For MizanQA and ArabCulture, benchmark-specific system prompts from the original papers are preserved.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#%F0%9F%8F%86-leaderboard-results🏆 Leaderboard Results

Results as of April 2026; covering top 10 evaluated models. Visit thelive leaderboardfor current rankings.

RankModelAVERAGEAraDiCE-CultureArabicMMLUArabCulturePALMX3LM STEMAraTrustMizanQAMedArabiQArabLegalQAGATMedAraBenchHumanEval+MBPP+FannOrFlop🥇 1Qwen/Qwen3.5-397B-A17B-FP868.06****82.7877.5461.7583.9188.6790.0473.3647.3054.9455.8947.9767.68****76.7244.33🥈 2Applied-Innovation-Center/Karnak66.2073.3380.9453.4981.4093.1089.0855.9255.7871.5861.0654.1933.5464.5558.91🥉 3inceptionai/Jais-2-70B-Chat65.8178.8981.29****83.2483.7387.9690.2371.7852.7969.6051.6750.8919.5143.6556.13#4Qwen/Qwen2.5-72B-Instruct65.7577.2273.7863.8377.7787.5588.5163.4950.0670.7455.9044.1937.2072.7557.51#5Applied-Innovation-Center/AIC-165.3773.3372.0277.5276.1188.1390.6156.3653.7568.9662.1150.7828.0569.5847.83#6Qwen/Qwen3.5-122B-A10B64.8474.4473.1737.7881.4686.1886.9764.0147.0455.1150.9052.4965.2472.4360.54#7Sakalti/Ultiima-72B64.4978.3372.2868.7976.7583.7089.0860.4444.5869.1246.9142.2539.0274.0757.56#8meta-llama/Llama-3.3-70B-Instruct63.9677.2271.5778.0577.9588.2885.6367.4456.2564.0051.1354.8627.4471.1624.43#9Qwen/Qwen2.5-32B-Instruct63.2670.5668.7675.8072.0781.0385.8253.7848.0869.2756.9436.5134.1572.7593.10#10FreedomIntelligence/AceGPT-v2-32B-Chat61.1476.6770.6279.7974.4684.8886.9763.8949.9671.4656.0447.3223.7854.5015.56

• **Scale does not guarantee best performance.**The top 10 spans models from 32B to 397B parameters, with several mid-size models outperforming larger ones on specific domains.
• **Arabic-specialized models lead on cultural and linguistic tasks.**Jais-2-70B-Chat ranks highest on ArabicMMLU and ArabCulture, while Karnak leads on 3LM STEM and ArabLegalQA.
• **Coding remains the hardest domain for Arabic-specialized models.**The top HumanEval+ and MBPP+ scores belong to multilingual models, with Qwen3.5-397B leading both.

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#the-size-performance-relationshipThe Size-Performance Relationship

Across the full leaderboard (46 models), a clear but imperfect size-performance correlation emerges. However, there are interesting exceptions:

• Arabic-specialized models often outperform size-matched multilingual models
• Instruction-tuned models consistently outperform their base counterparts except for Qwen3
• Some smaller Arabic-specialized models (Fanar-1-9B, ALLaM-7B) outperform much larger multilingual models on specific domains

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#%F0%9F%8C%9F-what-makes-qimma-different🌟 What Makes QIMMA Different

To summarize the distinctive properties of QIMMA:

PropertyDetailsQuality-first philosophyValidation runsbeforeevaluation, not as an afterthoughtMulti-model validationTwo LLMs with different training + human review for flagged cases99% native ArabicAvoids translation artifacts almost entirelyMulti-domain, multi-task7 domains, 3 task types (MCQ, QA, code), 109 subsetsCode evaluationFirst Arabic leaderboard to include code generationFull transparencyPer-sample inference outputs publicly released, not just aggregate scoresLightEval-basedUnified, reproducible evaluation codebaseDialectal awarenessExplicit handling of MSA vs. dialectal variation in prompts and rubrics

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#%F0%9F%94%97-resources🔗 Resources

• 🏆Leaderboard:QIMMA Leaderboard
• 💻Code:GitHub
• 📄Paper:Are Arabic Benchmarks Reliable? QIMMA’s Quality-First Approach to LLM Evaluation

https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard#%F0%9F%94%96-citation🔖 Citation

@misc{alqadi2026arabicbenchmarksreliableqimmas,
      title={Are Arabic Benchmarks Reliable? QIMMA's Quality-First Approach to LLM Evaluation}, 
      author={Leen AlQadi and Ahmed Alzubaidi and Mohammed Alyafeai and Hamza Alobeidli and Maitha Alhammadi and Shaikha Alsuwaidi and Omar Alkaabi and Basma El Amel Boussaha and Hakim Hacid},
      year={2026},
      eprint={2604.03395},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.03395}, 
}