Datasets:

Fatan-Cultural-Eval
/

fatan

Status: two rounds complete. The main evaluation round (5 models, 13 scenarios, 5 questions each, fully graded) and the NOTA round (correct-answer-removed MCQs, 5 models, 43 items, two turns each, fully graded) are both finished. A third, lighter round testing purely implicit recognition was considered and deliberately not run: the main round's open-ended versus multiple-choice contrast already evidences that axis directly, and a third full five-model pass was judged not worth the additional cost given what it would add. A neutral vision-only sanity check (confirming models can describe a scene correctly before being asked anything fiqh- or culture-loaded) was considered for the same reason and also set aside; its absence is recorded under Limitations rather than silently dropped.

1. The gap

Open-weight vision-language models marketed as "culturally aligned" or "Arabic-centric" are hypothesized here to not reliably demonstrate real Islamic/cultural competence, in a way standard VQA benchmarks rarely isolate, along two axes:

Implicit norm recognition vs. explicit Q&A. Does a model spot a cultural or fiqh-relevant detail unprompted, or only once a question has already named exactly what to look for and handed it a menu of plausible answers?
Specific identity recognition vs. flattening. Does a model distinguish Moroccan zellige from generic "Islamic art," or Amazigh Kabyle jewelry from generic "Arab jewelry," or does it collapse distinct regional and religious traditions into one broad label?

Most multimodal benchmarks for Arabic and Islamic content test general VQA accuracy or safety, not this specific explicit/implicit contrast, and rarely penalize identity-flattening at all.

Where this came from. This hypothesis came from my own hands-on, open-ended testing of Fanar and AIN before any of this pipeline existed. That testing is also why these two became the primary subjects here, rather than the broader field of Arabic VLMs: both make an explicit "culturally aligned" claim, and both failed repeatedly and confidently in that early testing. (Both also exceed the challenge brief's 6B parameter cap by about 1B; section 3 has the written permission I obtained for that from the Fatima Fellowship team, Abubakar Abid, before building the pipeline around them.)

In early exploratory sessions, I asked both models direct, plainly-worded, open-ended questions (not multiple-choice) about images of a dog on a prayer rug and a Gulf-style majlis: things like "is there anything here that conflicts with Islamic belief, given this is a Muslim household and the dog is lying on a prayer rug?" and "does this majlis scene conform to social customs?" Both models answered fluently and confidently. Neither consistently got the actual fiqh content right. AIN reached the right conclusion about the dog ("yes, there's a concern") but justified it with a fabricated mechanism, resting on a sacred object is forbidden, instead of the real one, najasa (ritual impurity). Asked separately about a mosque loudspeaker in another image, AIN invented a claim with no basis at all: that sound is a form of deity in Islam.

That gap between fluent, confident-sounding output and actual correctness, on exactly the content these models are marketed as being good at, is what this whole evaluation exists to measure systematically instead of anecdotally.

My role in this project. This was built as a close collaboration, not a case of handing an idea to an AI assistant and letting it run. A few concrete examples of decisions and catches that were mine, not Claude's: the two-axis framing above (implicit versus explicit, specific versus flattening) came from my own research background in alignment, before any code existed. I caught a real methodological flaw in an early version of the NOTA followup design, where models were shown their own prior wrong answer before being asked to reconsider, and called it out directly: that setup risks anchoring a model toward defending its own earlier answer rather than reasoning fresh, so the fix was a full two-turn regeneration, not a patch. I pushed for the LLM-judge grading to be audited rather than trusted, for both the NOTA round and the main round's checklist grading, instead of taking the automated verdicts at face value. I caught inaccuracies before they went into this document: a draft claim that the open-ended pass_checklist was human-written, when it wasn't; a Dataset Viewer config that would have made an unrun question set look like a completed round; and repeated instances of writing that read as generic AI phrasing rather than my own voice, which I asked to be rewritten plainly each time. I obtained written permission from the Fellowship organizers myself, by email, for the two primary models exceeding the stated parameter limit, before building the pipeline around them. And the scope decisions, not running a third implicit-recognition round, not running a vision-only control, were mine, weighed against what the main round's results already evidence, not defaults reached for by default.

2. Related work

Three bodies of existing work sit around this project, and it's worth being specific about what each one does and does not cover, since none of them tests quite what Fatan tests.

General-purpose vision-language hallucination benchmarks. POPE (Li et al., 2023) probes object hallucination through simple yes-or-no questions ("is there a {object} in the image?"), scored by recall, accuracy, and F1 across roughly 3,000 questions drawn from MSCOCO, A-OKVQA, and GQA. HallusionBench (Guan et al., 2024) goes further, pairing 346 human-annotated images with 1,129 questions specifically designed to separate visual illusion (the model misreads what's in front of it) from language-prior knowledge hallucination (the model's stored knowledge overrides what's actually shown). Both are general-purpose and object- or scene-level. Neither tests domain-specific factual competence (a fiqh ruling, a craft term, a named regional tradition), and neither includes a "none of the options is correct" condition the way Fatan's NOTA round does.

Truthfulness and imitative falsehood. TruthfulQA (Lin, Hilton, and Evans, 2022) introduced the framing most relevant to this whole project's premise: that larger, more capable models can perform worse on truthfulness specifically because their fluency lets them reproduce a common misconception more convincingly, not because they know less. Fatan's own findings (the fabricated Ibn Baz citation, the confident Fitr/Adha chronology reversal) are concrete instances of exactly this pattern, just in a visual, Arabic, religiously specific setting that TruthfulQA's English, text-only, general-knowledge questions don't cover. Bhatti and Alam (2025) document a more directly comparable pattern in Arabic specifically: converting Arabic MCQs into open-ended format across five dialects produces a substantial accuracy drop, with Arabic-centric models weakest on dialectal variants. Fatan's own main round finds the same shape of gap (MCQ accuracy 82-100% across all five models, open-ended checklist pass rate dropping to 55-82% on identical content), independently, on different content and a different model set, which is some evidence this is a real, general phenomenon rather than an artifact of one dataset's wording.

Cultural competence benchmarks, general and Arabic-specific. CulturalBench (Chiu et al., 2025; 1,696 human-written and human-verified questions across 45 regions) and BLEnD (Myung et al., 2024; 16 cultures, short-answer and multiple-choice) both document large, consistent gaps between well-represented and under-represented cultures, with BLEnD reporting a swing of over 50 percentage points between the best- and worst-served cultures on the same model. NormAd (Rao et al., 2025) specifically tests whether a model can judge social acceptability when the relevant norm must be inferred rather than stated outright, and finds performance drops sharply once the norm isn't handed to the model directly, a finding that resonates with Fatan's own explicit-versus-implicit axis. None of the three test Arabic or Islamic content specifically.

Closer to home, a small but growing set of Arabic-specific resources has emerged recently: Peacock (Alwajih et al., 2024), which pioneered Arabic multimodal VQA and OCR evaluation with a cultural-understanding component; Dallah (Alwajih, Bhatia, and Abdul-Mageed, 2024), which extends this to dialect-aware evaluation; CAMeL-Bench (Ghaboura et al., 2024), AraTraditions10K (Al-Buraihy et al., 2025), and Pearl (Alwajih, Magdy, et al., 2025), more recent additions to the same space; IslamicMMLU (Abdelaal et al., 2026), which tests Islamic knowledge specifically but in a text-only, standard-MCQ format; and PalmX 2025 (Alwajih, El Mekki, et al., 2025), the first shared task built specifically around Arabic and Islamic cultural benchmarking. As far as I've been able to determine from their published descriptions, none of these pairs an explicit multiple-choice question against an implicit, unprompted description task on the same scene the way Fatan's main round does, and none includes a NOTA-style, correct-answer-removed condition to test whether a model will admit it doesn't know rather than commit to the nearest plausible wrong option. That specific combination, explicit-versus-implicit plus epistemic humility under a removed correct answer, appears to be where Fatan sits that the existing literature doesn't already cover, though this dataset is far smaller than any of the benchmarks above and does not claim to replace them.

LLM-as-judge reliability. Zheng et al.'s MT-Bench and Chatbot Arena paper (Zheng et al., 2023) is the foundational reference for using a strong LLM as an automated grader, and the same paper that first documented the failure modes this project had to work around directly: position bias, verbosity bias, and self-enhancement bias, alongside the finding that a strong judge like GPT-4 can still reach over 80% agreement with human preference judgments in aggregate. Section 8's manual audit of all 29 "match" verdicts in the NOTA round, and the two specific error patterns it found (a missed obligatory-versus-conditional distinction, and category-level overlap mistaken for specific content), is a direct, practical instance of exactly the kind of judge unreliability this literature already predicts, not a problem unique to this project's setup.

Where this sits relative to my own work. This project is not an isolated exercise. My background is in preference-based post-training and alignment for Arabic-English foundation models, and I am a co-author on a separate, concurrent project, a cause-diagnostic Arabic visual-hallucination benchmark for multimodal LLMs, that tackles an adjacent question from a different angle: rather than testing whether a model can correctly describe or reason about a culturally specific scene, that project builds a taxonomy of why a model hallucinates on Arabic visual content in the first place (visual ambiguity, multimodal inconsistency, or incorrect cultural context, among its categories), using confusable captions and culturally close but wrong distractors rather than the open-ended and NOTA designs used here. The two projects share a research question, VLM reliability on Arabic and Islamic content, and diverge in method, which is itself useful: Fatan's finding that models default back to a closed-option frame even when explicitly told every option is wrong adds a concrete, model-behavior-level data point to the same broader problem that project is diagnosing at a cause level.

3. Models evaluated

Model	Role	Open-weight?	Params (approx.)
Fanar-2-Oryx-IVU (QCRI)	Primary subject, Arabic-first VLM, explicitly marketed as culturally aware	Yes	7B
AIN (MBZUAI)	Primary subject, Arabic-inclusive multimodal model	Yes	~7B
Qwen3-VL-4B-Instruct	Generic multilingual baseline	Yes	4B
Gemma-3-4b-it	Generic multilingual baseline	Yes	4B
GPT-5 (via OpenRouter)	Closed-model reference point only, not the graded subject	No	N/A

Fanar and AIN are the primary subjects per the challenge brief (a frontier-for-its-class, open-weight model you can load yourself). Both are tested side by side since they're the same parameter class and make the same "culturally aligned" claim. Qwen3-VL and Gemma contextualize whether any Fanar/AIN advantage is real or illusory. GPT-5 is a closed reference point, included for context, not evaluated against the brief's open-weight requirement.

Qwen3-VL-4B-Instruct, specifically, was picked over an older Qwen generation on purpose, not by oversight: it was released May 14, 2025, close to Gemma-3-4b-it's March 12, 2025, keeping the two generic baselines contemporaries of each other rather than pairing a newer baseline against an older one. Matching generations matters here more than matching exact release dates would for most comparisons, since a newer generic baseline is, if anything, a harder bar for Fanar and AIN to fail to clear, not an easier one.

A note on the 6B parameter limit. The challenge brief caps open-weight models at 6B parameters. Fanar-2-Oryx-IVU (7B) and AIN (~7B) both exceed that by roughly 1B, since they were, by a clear margin, the models that best fit the specific multimodal and Arabic-cultural question this project investigates; no compliant Arabic-first VLM in the 6B-and-under range covers the same ground. I emailed the Fatima Fellowship team about this before building the pipeline around these two models and received written approval to proceed:

"Yes that is fine. That works for us, please include a note confirming that you emailed us about this." Abubakar Abid, Fatima Fellowship, September 29, 2026, in reply to my email asking whether a 2-3B deviation above the stated limit would be acceptable if justified and documented.

4. Dataset

13 AI-generated scenario images (7 Islamic-fiqh, 6 cultural-heritage), each built with a human-in-the-loop pipeline: Gemini generated images and draft questions, ChatGPT cross-checked them, and every factual claim was then manually verified against primary sources (binbaz.org.sa, islamqa.info, UNESCO Intangible Cultural Heritage listings, national tourism and culture ministries, and museum collections including the Met and the Smithsonian). Several citation errors and one internal contradiction were caught and fixed during this verification pass (see commit history on fatan_scinarios.json).

Each scenario carries 3 MCQs plus 2 open-ended questions, one analytical and one counterfactual, each graded against a pass_checklist of the specific points a correct answer should make. The checklist went through the same process as the rest of the dataset: Gemini drafted it, ChatGPT cross-checked it, and it was manually verified against the same primary sources used throughout (binbaz.org.sa, islamqa.info, UNESCO ICH listings, and others), not written from scratch by hand.

The two open-ended types are curated differently, not just worded differently. Every analytical question (13 of 13) opens by describing the specific scene in the image ("تُظهر الصورة...") before asking how a stated fiqh principle applies to it, so answering it well requires engaging with the particular scenario shown, not just the general rule. The counterfactual question is a "what if" variation on that same scenario, meant to test whether a model's reasoning tracks a changed fact rather than a memorized answer, but it's less consistently tied back to the image: only 6 of 13 (46%) explicitly reference it, since some counterfactuals change a detail of the depicted scene (what if she hadn't known the earrings had a cross on them?) while others pose a more general hypothetical not anchored to anything visible (what if the prayer were on open ground instead of a carpeted mosque?). This is a different curation approach from the NOTA round below, which doesn't author new open-ended prompts at all; it reuses the existing MCQ stem and strips the correct option out.

A second question set, fatan_nota_questions.json, was built on top of the same 13 images for the NOTA round (section 7). Each of its 43 items takes one MCQ from the original set and removes the correct option, replacing it with a literal "none of the above" choice, so that none of the four visible options is actually correct. This set went through the same verification discipline as the main dataset and caught real problems along the way: an early draft of several items leaned on AI-generated claims about craft terminology, regional dress, and material history that turned out on checking to be wrong or unverifiable (a fabricated Tunisian embroidery school name, an incorrect attribution for the Moroccan riad fountain term, among others) and were corrected or dropped before the items were used. One item in the final set (the traditional zellige-cutting tool, correctly verified as al-manqash) and several of the distractor sets for the Omani khanjar and Moroccan zellige items were specifically checked against primary sources rather than accepted on first draft.

Full dataset, images, pipeline code, and raw results: https://huggingface.co/datasets/Fatan-Cultural-Eval/fatan

A note on where this lives: development happened in Fatan-Cultural-Eval/fatan-eval, which accumulated 80 commits across the iteration described throughout this document. That repository has been kept private, and its final state was migrated as a single squashed commit into the public repository linked above for submission.

5. Method

The three rounds form a deliberate progression, not three unrelated tests. Each strips away one more piece of scaffolding a model could lean on without actually knowing the material: the main round's MCQs present the correct answer among four visible options; the main round's open-ended questions ask the same content with no options at all; and the NOTA round's MCQs remove the correct option entirely and say so outright. This progression is grounded in two findings from outside this project worth citing directly, not just intuition: NormAd (Rao et al., 2025) finds that a model's accuracy on judging social norms drops sharply once the norm must be inferred rather than stated outright, evidence that recognition and production are genuinely different skills, not two measurements of the same thing; and Bhatti and Alam (2025) find, in Arabic specifically, that converting MCQs to open-ended format produces a substantial accuracy drop, with Arabic-centric models weakest on dialectal variants. The expectation going in was that if "culturally aligned" claims were real, that gap should be small for Fanar and AIN specifically; sections 6 through 8 report what was actually found.

5.1 Main round

For every model, every scenario, the identical fixed question sequence (mcq_1 -> mcq_2 -> mcq_3 -> open_a -> open_b) runs as one continuous conversation, no adaptive branching, same order and wording for every model. MCQs are auto-scored by extracting the chosen letter (regex, with an LLM fallback for messy answers) and comparing it to the known-correct option. Open-ended answers are graded by Gemini acting as an LLM-as-judge: for each pass_checklist item, a true/false judgment with a one-sentence rationale. Full pipeline code (fatan_eval_pipeline.py / .ipynb) is in the dataset repo above.

5.2 NOTA round

Each of the 43 NOTA items runs as its own short, independent two-turn conversation per model (not chained to the other NOTA items for the same image, so a wrong answer on one item cannot prime the next):

Initial turn. The model is shown the image and the four-option MCQ, one of which literally reads "there is no correct answer among the options given," and is instructed to answer with a single letter and no explanation.
Followup turn. In the same conversation, the model is then asked: "So, in a complete sentence, what do you think the correct answer is, and why? Explain your reasoning, don't just give a single letter."

This produces two independent signals per item: whether the model recognizes a genuine absence of a correct option on the first try, and whether it can actually produce the real answer when pushed on the second. The real answer never appears anywhere in what the model is shown; it exists only as a hidden reference used for grading.

The followup is graded separately, by Gemini comparing the model's free-text response against the hidden reference answer and returning a binary match judgment with a one-sentence rationale, the same LLM-as-judge pattern used for the main round's open-ended questions but adapted to a single reference rather than a checklist.

This design went through two significant revisions after problems surfaced in the data itself, both described in section 9.

6. Main round results

325/325 rows complete and graded (5 models, 13 scenarios, 5 questions each).

MCQ accuracy

Model	Accuracy
GPT-5 (closed, reference only)	100.0% (39/39)
Qwen3-VL-4B-Instruct	92.3% (36/39)
Fanar-2-Oryx-IVU	87.2% (34/39)
AIN	82.1% (32/39)
Gemma-3-4b-it	82.1% (32/39)

On explicit multiple-choice alone, Fanar and AIN do not clearly outperform the generic multilingual baselines. Qwen3-VL actually scores higher than both, and AIN ties the lowest score with Gemma.

Open-ended checklist pass rate (the sharper signal)

Model	Checklist items satisfied
Gemma-3-4b-it	82.0%
GPT-5 (closed, reference only)	80.9%
Fanar-2-Oryx-IVU	61.8%
Qwen3-VL-4B-Instruct	61.8%
AIN	55.1%

This is the more revealing result. MCQ accuracy rewards recognizing a correct option among four; the open-ended checklist rewards actually producing sound, complete reasoning unprompted by options. Here, both "culturally aligned" models fall to the bottom of the field, well below both generic multilingual baselines. A model can apparently recognize the right fiqh ruling when it's handed to it as one of four choices, without being able to reconstruct the full reasoning behind it unprompted, exactly the explicit-vs-implicit gap this evaluation set out to measure.

By domain (Islamic fiqh vs. cultural heritage), MCQ only

Model	Fiqh	Cultural
Fanar-2-Oryx-IVU	95% (20/21)	78% (14/18)
AIN	81% (17/21)	83% (15/18)
Gemma-3-4b-it	86% (18/21)	78% (14/18)
Qwen3-VL-4B-Instruct	90% (19/21)	94% (17/18)
GPT-5	100%	100%

Fanar's single strongest result (95% on fiqh MCQs) is also its weakest domain split (78% cultural), consistent with Fanar's training emphasis being religious and linguistic rather than broader regional cultural identity, though 18 to 21 items per cell is too small to treat this as more than a directional signal.

7. NOTA round results

430/430 rows complete and graded (5 models, 43 items, two turns each, zero outstanding errors).

Core results

Model	Recognized the trap (picked "none of these")	Actually knew the real answer
GPT-5 (closed, reference only)	90.7% (39/43)	46.5% (20/43)
Qwen3-VL-4B-Instruct	20.9% (9/43)	2.3% (1/43)
Fanar-2-Oryx-IVU	18.6% (8/43)	7.0% (3/43)
AIN	14.0% (6/43)	4.7% (2/43)
Gemma-3-4b-it	7.0% (3/43)	2.3% (1/43)

The "actually knew the real answer" column reflects a correction to the automated grading, described in section 8: three matches that the LLM-judge initially marked correct did not hold up on manual review and were reclassified, affecting AIN (down from 4/43 to 2/43) and Qwen3-VL (down from 2/43 to 1/43). GPT-5's count moved the other way, from 19/43 to 20/43, after a full audit of every "no match" verdict (also section 8) turned up one case the judge had wrongly rejected on formatting grounds rather than content. Fanar's and Gemma's numbers were unaffected.

With four options and one designed to be the honest answer, random guessing on the first column alone would land near 25%. Three of the four open-weight models sit at or below that baseline. None of them comes close to demonstrating that it actually knows the real answer once the multiple-choice crutch is taken away: the best open-weight result is 7.0%, against GPT-5's 46.5%.

The deeper pattern: models cannot leave the four-option frame

The followup question is explicitly open-ended and explicitly tells the model that none of the original four options were correct. Despite that, most models answer it by picking one of those same four letters anyway:

Model	Cites a specific letter as "the answer" in the followup	Of those, cites one of the three already-rejected letters
Gemma-3-4b-it	100% (43/43)	97.7% (42/43)
Qwen3-VL-4B-Instruct	97.7% (42/43)	76.7% (33/43)
Fanar-2-Oryx-IVU	90.7% (39/43)	72.1% (31/43)
AIN	90.7% (39/43)	81.4% (35/43)
GPT-5	4.7% (2/43)	0% (0/43)

This is, in my read, the single sharpest finding in the NOTA round. Four of the five models cannot escape the closed-option frame even when directly instructed to and even when told outright that every option is wrong. Gemma does this on 42 of its 43 followup answers: it states a specific letter as "the correct answer" in bold, then writes a paragraph of justification for an option that was, by the design of the question, already proven false. GPT-5 is the only model that reliably reasons in the open space it's actually being asked to reason in.

This also explains why an earlier version of this dataset showed a 0% followup-match rate for Fanar, AIN, and Qwen3-VL: in that pass, all three simply repeated a single letter in response to the followup question, apparently carrying the first turn's "answer with only a letter" instruction over into a differently worded second question that never asked for one. The followup prompt was rewritten to explicitly request a full-sentence answer with reasoning, and all three models were rerun from a completely fresh two-turn conversation (not a replay of the old letter plus a new followup, since replaying a model's own earlier answer into context before asking it to reconsider risks anchoring it toward defending that answer rather than reconsidering the question). The rerun fixed the literal one-character responses. It did not fix the underlying behavior: the models still anchor on one of the four options, they just now write confident, fabricated justification around that choice instead of staying silent about it. That shift from a formatting bug to a content problem is itself informative.

8. Verifying the grader

Because the followup turn's accuracy depends entirely on an LLM judge's comparison between free text and a hidden reference, every one of the 29 cases across all five models where Gemini marked a followup as matching the real answer was checked by hand against the full raw model response, not the summary, before being reported above. Twenty-six held up as genuinely correct, including subtler cases such as GPT-5 citing a different but still factually correct supporting Quranic verse than the one in the reference. Three did not, and they fall into two distinct, specific failure patterns rather than one general unreliability:

Obligatory versus conditional. AIN was asked about Saudi shoe-removal custom and answered that removing shoes is "wajib" (obligatory, universal). The reference answer's entire point is that it's conditional on the household's custom or the host's direction, not an absolute rule. That conditional-versus-absolute distinction was the exact thing separating the correct answer from a rejected wrong option in the original question. The judge checked only whether shoe removal was mentioned and missed that the model had actually restated the rejected framing.
Category overlap mistaken for specific content. AIN was asked about the meaning of Palestinian tatreez patterns. The reference requires three specific claims: regional identity tied to named cities (Hebron, Jaffa, Ramallah), the wearer's marital or social status, and the occasions the dress is worn. The model's answer gestured broadly at "cultural, social, economic, political, symbolic" meaning without stating any of the three specific claims. The judge accepted the broad topical overlap as if it satisfied the specific content requirement.
Judge comprehension failure on degenerate text. Qwen3-VL's response to a question about wearing a cross necklace argued the opposite of the correct ruling (that the cross does not need to be removed, when the reference requires removal or concealment), then degenerated into the same clause repeated roughly thirty times before hitting the token limit. The judge's rationale claimed the model had "advised avoiding wearing symbols like the cross," which the model never actually said anywhere in the response. This looks like the judge losing track of the actual content once the text became a long, confusing wall of repetition, rather than a rubric problem.

The corrected figures in section 7 reflect this audit. The original, unverified judge output was not deleted; both the automated verdict and the manual correction are preserved in fatan_nota_results.json for anyone who wants to check the reasoning independently.

The much larger set, the 186 followup cases Gemini marked as "does not match," was also fully audited by hand, not spot-checked. Every case was triaged by lexical overlap between the raw response and the hidden reference (the cases most likely to hide a missed match are the ones where the response and reference share the most vocabulary despite the judge calling them different), then read in that order from highest to lowest overlap, with the full raw response checked against the judge's own stated rationale.

185 of the 186 held up as genuinely correct rejections. One did not: GPT-5's response to the shoe-impurity question in image_b70c76 substantively restates the correct ruling (wipe or rub the sole on clean ground until the impurity is gone, no water required, then continue praying), which matches the reference in content. The judge's own rationale rejected it for not stating the answer "in the concise manner of the reference answer" and not translating the hadith, not because the content actually disagreed. That is the opposite failure mode from the three found above: those were the judge being too lenient or losing track of degenerate text; this one is the judge being too strict about phrasing when the substance was right. Worth noting because it means the judge's errors in this dataset don't all point the same direction, which is itself evidence against a simple one-sided bias in the grading.

The main round's checklist grading got the same treatment. All 445 checklist judgments behind section 6's open-ended pass rates (130 question-rows: 13 scenarios x 2 open-ended questions x 5 models) were checked by hand against the raw responses, not sampled. Zero fabrication errors turned up anywhere in the full set: no case of the judge crediting a response with content it never actually contained, and no case of a genuine match wrongly rejected, the two failure types that mattered in the NOTA audit above are simply absent here at full coverage.

What did recur, narrowly, is a softer pattern: 7 of 304 "met: true" verdicts (2.3%) credit a response that only partially satisfies a multi-part checklist item, sometimes while the judge's own rationale admits the gap. Two representative cases: asked why shoes are avoided on mosque carpets specifically, AIN's response only says open ground "isn't sacred like a mosque," never stating the actual carpet-cleanliness and dispute-avoidance reasoning the checklist item requires, yet is marked met. Asked about the Moroccan riad's architectural indicators, both AIN's and Qwen3-VL's responses mention only the fountain and omit the plants, and in one case the judge's own rationale explicitly notes "doesn't mention plant elements" before marking the item true anyway. This pattern is spread across 4 of the 5 models rather than concentrated on one, and does not change any model's ranking or published number: the figures in section 6 (Gemma 82.0%, GPT-5 80.9%, Fanar 61.8%, Qwen3-VL 61.8%, AIN 55.1%) stand as reported.

One genuine model-behavior finding surfaced during this audit, not a grading error: on the cross-necklace scenario (image_crx005), AIN, Fanar, and Gemma each independently claimed that prayer performed while unknowingly wearing cross-jewelry is invalid and must be repeated, the opposite of the actual ruling (valid, no repetition needed, since ignorance of the symbol exempts the person from blame). The judge caught this correctly in all three cases. Three separate models converging independently on the same specific wrong fiqh point is a notable behavioral data point in its own right; see section 9.

9. Behavioral patterns worth naming directly

Pull the three tiers from section 5 together and a single pattern emerges. When the correct answer sits among four visible options, every model scores well (82-100% MCQ accuracy, section 6). Strip the options and ask the same content open-ended, and the two "culturally aligned" models, Fanar and AIN, fall to the bottom of the field, below both generic baselines (55.1% and 61.8%, against 61.8-82.0% for the others, section 6). Remove the correct option entirely and tell a model outright that none of the four choices is right, and most models still cannot let go of the frame: 72.1-97.7% of them cite one of the three already-rejected options anyway when asked to reason freely, and only answer with real, grounded content in a small minority of cases, 2.3% to 7.0% for the open-weight models, against GPT-5's 46.5% (section 7). The more scaffolding is removed, the less these models actually know, and the more confidently they fabricate around the gap, including inventing citations to named religious authorities that contradict what those authorities actually hold. The examples below are concrete instances of that same pattern, kept because they are more useful for understanding what is actually going wrong than the aggregate numbers alone.

Fabricated scholarly attribution. Asked about praying with shoes on in a carpeted mosque, Fanar restated a rejected wrong option and justified it by attributing a specific, confident position to Sheikh Ibn Baz: that the practice is impermissible except under necessity. The actual, verified correct answer (and Ibn Baz's real documented position) is the opposite: praying in clean shoes is sunnah and recommended, grounded in the Prophet's own practice. The model did not just get the ruling wrong; it invented a citation to a real, named religious authority that contradicts what that authority actually holds. This is not a one-off: on a separate item about the two Eid holidays, AIN fabricates a specific nonexistent book title, "الإمام الشافعي وعقيدته," and attributes it to Ibn Baz. Different model, different question, same underlying behavior: inventing a precise, checkable citation to a real authority rather than saying "I don't know" or hedging.

Self-contradiction within a single response. On a different item in the same domain, Fanar stated in one sentence that praying with shoes in a carpeted mosque is forbidden, then in the next sentence instructed the reader to please wear shoes upon entering, two directly opposed claims inside the same short paragraph.

A basic chronology error stated with full confidence. Asked about the two legitimate Eid holidays, Fanar justified a wrong answer by claiming Eid al-Fitr is "the first holiday Muslims celebrate after Eid al-Adha," which reverses their actual order in the Islamic calendar. This is not a subtle fiqh distinction; it is a basic, checkable fact stated backwards with the same confident tone as everything else in the response.

Three independent models converging on the same wrong ruling. Asked about prayer performed while unknowingly wearing cross-shaped jewelry, AIN, Fanar, and Gemma each independently claimed the prayer is invalid and must be repeated. The actual ruling is the opposite: the prayer is valid and does not need to be repeated, since not knowing the jewelry bore a cross exempts the person from blame. This surfaced during the main round's checklist-grading audit (section 8), not as a grading error, the judge correctly marked all three responses as failing that checklist item. What makes it worth naming on its own is that three separate models, with no shared prompt artifact or carried-over context, converged independently on the identical specific wrong answer, which looks less like three unrelated mistakes and more like a shared gap in how this particular ruling is represented across training data.

Cross-script leakage during generation. Two separate instances, in two different models, of a foreign, non-Arabic character appearing mid-word inside otherwise Arabic text: Fanar's response to the Eid question contains the literal fragment "هذا因为," splicing a Chinese character meaning "because" into the middle of an Arabic sentence. Gemma's response to a shoes-in-mosque question contains "الأچارب," which embeds a Persian or Urdu letter not present in the standard Arabic alphabet, producing a word that does not actually exist. The same "هذا因为" glitch recurs in a second, separate Fanar response (on a question about Moroccan zellige tools), confirming it is a recurring artifact of generation under this setup rather than a one-time fluke. Both occurred inside otherwise fluent-sounding Arabic prose, which is precisely what makes them easy to miss on a casual read and exactly the kind of confident, fluent-but-wrong output this whole evaluation is built to surface rather than overlook.

I want to flag one of these patterns as corroborated outside this project rather than a one-off artifact of this specific prompting setup. I encountered the same two behaviors, degenerate repetition loops and unexpected non-Arabic character leakage, in Qwen-family models during my own earlier work on Saudi dialect and Modern Standard Arabic alignment at KAUST, working with both dialectal and standard registers. Seeing it again here, independently, in a completely different evaluation and a different Qwen-based model, is evidence that this is a systematic weakness in how this model family handles Arabic generation under pressure, not something specific to how these particular questions were phrased.

Genuine same-conversation contradiction. Two clean cases survive in the final dataset of a model reversing its own answer between the initial and followup turns of the same conversation, with no new information given in between, both from Gemma. On the cross-necklace item in image_crx004 (mcq_2), Gemma's initial turn correctly selects C, "none of the options is correct." On the followup turn it reverses to D and explicitly disowns its own prior answer: "الخيار C غير صحيح" ("option C is incorrect"), contradicting what it had said one turn earlier in the same conversation. The same pattern recurs on a separate item about Moroccan zellige wall decoration (image_clt005, mcq_2): correct "C, none of the options" on the initial turn, reversed to a specific wrong option, B, on the followup.

No new evidence was introduced between these turns. The only thing that changed was being asked the same underlying question a second time. That is a direct, data-grounded instance of the "fluent but not grounded" pattern this project set out to measure, not an inference from aggregate accuracy numbers.

10. Issues encountered building the NOTA round

Documented here because several of them are informative in their own right, not just engineering friction.

GPT-5's reasoning budget. GPT-5 can spend its entire token budget on hidden reasoning before producing any visible answer, returning an empty response. This was first hit and fixed in the main pipeline (the open-ended question budget was raised from 400 to 1000 tokens). It resurfaced in the NOTA round in a new place: the initial MCQ turn only needs a handful of tokens for an ordinary model to output a single letter, but a NOTA item specifically requires reasoning through why each of four plausible-looking options is wrong before committing to "none of these," which used measurably more of GPT-5's hidden reasoning budget than a normal MCQ does. The same 200-token budget that worked without issue for the main round's ordinary MCQs produced the same empty-response failure here. The fix (raising the budget to 600 for the initial turn and 1600 for the followup) costs nothing for a model that stops early on its own; it only matters for the one model that actually needs the room, which is itself a small piece of evidence that these items demand more real reasoning than a standard MCQ.

A GPU memory failure mid-run. Fanar hit a CUDA out-of-memory error partway through a long run of back-to-back generation calls, a known fragmentation pattern on a memory-constrained free-tier GPU rather than any single call being too large. The pipeline's existing per-call error handling caught it, logged the failure, and continued to the next item instead of crashing the whole run; the failed item was retried automatically on the next pass.

A silent data-overwrite bug. The function responsible for downloading the NOTA question set tried the Hub first and only fell back to a manually uploaded local copy if that request failed outright. A stale, pre-correction version of the question file already existed on the Hub, so the Hub request succeeded and silently overwrote a corrected local upload before a single model had loaded, with no error raised and no indication anything had gone wrong. An entire round of fact-checking work was quietly discarded this way before it was caught. The fix makes a locally present file take priority unconditionally, with the Hub used only when nothing local exists at all.

The single-letter followup bug and why the first fix wasn't enough. Described in section 7. Worth restating the sequence here because it reflects a real methodological lesson: rewording the followup prompt alone did not fix the underlying issue, because the conversation still carried the model's own earlier letter choice in its context before asking it to reconsider, which risks anchoring the model toward justifying that earlier answer rather than genuinely reconsidering the question. The actual fix required regenerating both turns of the conversation from a clean state for the affected models.

A missing-package failure after that fix. Once the fresh two-turn rerun was deployed, Fanar and AIN failed on all 43 of their items with ModuleNotFoundError: No module named 'qwen_vl_utils', a package needed only by the two Qwen2/2.5-VL-based chat code paths and not by Qwen3-VL or Gemma, which is why those two succeeded while Fanar and AIN failed completely. The cause traced back to a partial, ad hoc installation: an earlier bitsandbytes upgrade, run to fix an unrelated model-loading error, was mistaken for having also installed the separate package needed for image processing during generation. Resolved once the full install step was run in the same kernel session, before inference began.

A gap in the resumability design. The pipeline could fetch the NOTA question set automatically from the Hub but had no equivalent step for fetching prior results, so a genuinely fresh runtime with nothing manually uploaded would see zero completed rows and redo already-finished work, including models that had already been fully graded. Fixed by adding the same "pull from the Hub only if nothing exists locally" logic already used for the question set.

Free-tier disk pressure. Caching four separate open-weight models' downloaded parameters across a multi-day iteration process eventually filled the runtime's disk. Addressed by distinguishing a full runtime reset, which gives a clean disk and clean memory but requires re-downloading everything, from a simple kernel restart, which keeps the same disk and therefore does not actually free any space.

11. Proposed path forward (speculative)

My own background is in preference-based post-training and alignment for Arabic-English foundation models (most recently as a Visiting Researcher at KAUST). That's not incidental to where these proposals land: each one below is something I'd know how to actually build, and each is aimed at a specific failure pattern observed in this project's own data or in my earlier exploratory testing of Fanar and AIN, not a generic "fine-tune on more data" suggestion.

Fine-tuning paradigm: domain-triggered extended reasoning. Detect when a query touches religious or cultural content and route it to an extended-reasoning mode before generation, rather than the same fast-path used for any other question. This targets AIN's dog/prayer-rug answer directly: it reached the right conclusion ("yes, there's a concern") but justified it with a fabricated mechanism ("resting on a sacred object is forbidden") instead of the real one (najasa, ritual impurity). It is also directly supported by the NOTA round's letter-persistence finding: a model trained to pause and reconstruct the reasoning chain for sensitive domains, rather than default back to the nearest multiple-choice-shaped answer, is less likely to produce a confident answer built on an option it was already told was wrong.
Data curation and architecture: hallucination becomes explicit refusal with referral, not a confident guess. When confidence is low, the model should say so directly and point to where a human could verify, the same category of sources this project used to fact-check its own dataset (binbaz.org.sa, islamqa.info, UNESCO ICH listings), rather than answering anyway. This responds to AIN's loudspeaker hallucination (confidently inventing "sound is a form of deity in Islam" out of nothing) and to the NOTA round's fabricated Ibn Baz attribution: in both cases the fix isn't more correct facts in training data, it's training the model to recognize when it is extrapolating past what it actually knows and to say so, with somewhere real to send the user instead of a fabricated answer or an invented citation.
Architecture: retrieval before answering, with the output explicitly framed as reasoning, not a ruling. Pair retrieval against a vetted corpus (comparable to the sources used to build this dataset) with an output format that distinguishes "here is my reasoning, grounded in these sources" from an authoritative answer. This matters even when retrieval succeeds: for religious rulings specifically, an AI asserting a definitive fatwa is a problem of claimed authority, not just accuracy. The model should be able to say "based on these sources, it appears that..." rather than issuing a ruling outright. It's also informed directly by a problem hit while building this dataset: Gemini repeatedly cited the wrong fatwa page when generating scenario references, caught only by manually re-verifying every citation against the source and cross-checking with ChatGPT. The dataset-construction process ran into the same failure mode the dataset now measures.
Fine-tuning paradigm: preference pairs rated by domain experts, not crowdworkers. The flattening-of-distinct-traditions axis (Moroccan zellige read as generic "Islamic art," Kabyle jewelry read as generic "Arab jewelry") is a preference-alignment problem as much as a knowledge one: it's about what a response is rewarded for during RLHF, not just what facts are available to it. Using scholars and regional-heritage specialists as preference raters, rather than generalist annotators, would shift the reward signal toward distinguishing specific traditions instead of toward fluent, broadly-plausible-sounding Arabic.

None of these are validated here. They're reasonable next steps given what the gap looks like, not proven fixes.

12. Limitations

The dataset-construction process hit the exact failure mode it studies, twice. Gemini, used to draft main-round scenario references, repeatedly cited the wrong fatwa page; every citation in that dataset was manually re-verified against the actual source and cross-checked with ChatGPT before being trusted. The NOTA round's early drafts carried the same problem in a different shape: specific claims about craft terminology and regional material culture that read as confident and plausible but turned out, on checking against primary sources, to be wrong or unverifiable, and had to be corrected or dropped before use. Worth stating plainly rather than glossing over: confident, wrong output wasn't just something the evaluated models did.
The LLM-judge grading was verified, not assumed, and still has a known, narrow error rate in both directions. All 29 "followup matches" verdicts were checked by hand; three did not hold up, split across two distinct failure patterns described in section 8. The full set of 186 "does not match" verdicts was also checked by hand, not just spot-checked; one did not hold up, for a third, distinct failure pattern (the judge penalizing phrasing rather than checking content). The corrected figures in section 7 account for all four. Net error rate: 4 out of 215 followup verdicts, roughly 10% within the much smaller "matched" set against roughly 0.5% within the much larger "did not match" set, which suggests the judge leans conservative overall rather than erring evenly in both directions. The main round's checklist grading (section 6) received the same full audit, all 445 judgments behind it, not a sample, and held up better: zero fabrication-level errors, against a narrow 2.3% soft-positive pattern (7 of 304 "met" verdicts crediting a partially-satisfied multi-part item), detailed in section 8. None of it changes the published per-model pass rates.
Most of the MCQ round cannot distinguish visual grounding from text-only knowledge. Only 20% of main-round MCQ questions (8 of 40) and 26% of NOTA questions (11 of 43) explicitly reference the attached image in their text ("تُظهر الصورة..."); the rest are phrased as free-standing fiqh or cultural-knowledge questions that happen to have an image attached, not ones the model is textually directed to look at. For that majority, a model could in principle produce the identical answer with the image removed entirely, since nothing in the question forces it to use what it sees. The open-ended questions are the exception: 73% (19 of 26) explicitly reference the image, and the handedness and hand-position items (which hand holds the dallah, which hand takes the cup) are a genuine positive control, since they are unanswerable without actually looking. No neutral "describe this image" control and no text-only, image-removed condition were run to measure this directly, so the main round's MCQ accuracy numbers should be read as evidence of cultural knowledge first and multimodal grounding second, with the open-ended round and the handedness items carrying the stronger visual-grounding evidence.
A third round testing purely implicit recognition, independent of the open-ended versus multiple-choice contrast already in the main round, was not run. This was a deliberate scope decision rather than an oversight, made after weighing the additional evidence it would add against the cost of a third full five-model pass, given that the main round's existing open-ended results already evidence the same axis directly.
Cultural coverage skews toward what I'm personally familiar with. The six cultural-heritage scenarios span Saudi, Omani, Palestinian, Moroccan, and Amazigh/Kabyle content, researched rather than lived. Six scenarios cannot claim to represent the Arab world's cultural diversity, and I don't want to imply broader coverage than this dataset actually has.
Built entirely on free-tier Google Colab, which capped model size (4B to 7B, not larger) and repeatedly interrupted iteration through GPU-hour limits, runtime disconnects, missing-package failures after a VM reset, and disk-space pressure from caching multiple models' weights across a multi-day process. The pipeline's resumability logic exists specifically because of how often this happened during development, a reasonable engineering response to a real constraint, not a hidden one. Several of the specific failures and fixes are documented in section 10 rather than smoothed over.
Generation was deterministic (greedy decoding), not sampled. Every result here reflects a single run per model per item. No variance estimate from repeated sampling exists, so a borderline result (for example, the exact gap between Fanar's and AIN's NOTA-correct rates) should be read as directional rather than statistically precise.

13. Repository contents

fatan_scinarios.json: the 13 main-round scenarios (images, questions, checklists, references)
fatan_nota_questions.json: the 43-item NOTA round question set, fact-checked and finalized
images/: the 13 scenario images, shared by both rounds
fatan_eval_pipeline.py / .ipynb: the main-round inference and grading pipeline (Colab-ready)
fatan_nota_eval_pipeline.py / .ipynb: the NOTA round's separate inference and grading pipeline (Colab-ready, independently resumable)
fatan_results.json: raw, fully-graded main-round results (325 rows)
fatan_nota_results.json: raw, fully-graded NOTA round results (430 rows), including both the original automated judge verdicts and the manual corrections described in section 8
fatan_implicit_questions.json: a drafted implicit-recognition question set, built but deliberately not run (see Limitations)
proposal/Research_Proposal.pdf / PROPOSAL.md: the full proposed-path-forward document (section 11), as its own standalone file

14. References

Abdelaal, A., Al Haffar, M. N., Fawzi, M., & Magdy, W. (2026). IslamicMMLU: A benchmark for evaluating LLMs on Islamic knowledge. arXiv:2603.23750.

Al-Buraihy, E., et al. (2025). AraTraditions10k: Bridging cultures with a comprehensive dataset for enhanced cross-lingual image annotation, retrieval and tagging. Scientific Reports. (Full co-author list not independently verifiable past the lead author; the underlying venue and paper are confirmed via a published correction notice.)

Alwajih, F., Bhatia, G., & Abdul-Mageed, M. (2024). Dallah: A dialect-aware multimodal large language model for Arabic. arXiv:2407.18129.

Alwajih, F., El Mekki, A., Mubarak, H., Hawasly, M., Mohamed, A., & Abdul-Mageed, M. (2025). PalmX 2025: The first shared task on benchmarking LLMs on Arabic and Islamic culture. arXiv:2509.02550.

Alwajih, F., Magdy, S. M., El Mekki, A., Nacar, O., Nafea, Y., Abdelfadil, S. T., Yahya, A. M., Luqman, H., Almarwani, N., Aloufi, S., Qawasmeh, B., Atou, H., Sibaee, S., Alsayadi, H. A., Al-Dhabyani, W., Al-shaibani, M. S., El Aatar, A., Qandos, N., Alhamouri, R., Ahmad, S., Al-Ghrawi, M. A., Yacoub, A., AbuHweidi, R., Mohamed Lemin, V., Abdel-Salam, R., Bashiti, A., Alansari, A., Ashraf, A., Alturayeif, N., Alcoba Inciarte, A., Ammar, A., Elmadany, A. A., Tourad, M. C., Berrada, I., Jarrar, M., Shehata, S., & Abdul-Mageed, M. (2025). Pearl: A multimodal culturally-aware Arabic instruction dataset. arXiv:2505.21979.

Alwajih, F., Nagoudi, E. M. B., Bhatia, G., Mohamed, A., & Abdul-Mageed, M. (2024). Peacock: A family of Arabic multimodal large language models and benchmarks. ACL 2024 (per ACL Anthology 2024.acl-long.689). arXiv:2403.01031.

Bhatti, H. H., & Alam, F. (2025). Beyond MCQ: An open-ended Arabic cultural QA benchmark with dialect variants. arXiv:2510.24328.

Chiu, Y. Y., Jiang, L., Lin, B. Y., Park, C. Y., Li, S. S., Ravi, S., Bhatia, M., Antoniak, M., Tsvetkov, Y., Shwartz, V., & Choi, Y. (2025). CulturalBench: A robust, diverse, and challenging cultural benchmark by human-AI CulturalTeaming. ACL 2025. arXiv:2410.02677.

Ghaboura, S., Heakl, A., Thawakar, O., Alharthi, A., Riahi, I., Saif, A., Laaksonen, J., Khan, F. S., Khan, S., & Anwer, R. M. (2024). CAMEL-Bench: A comprehensive Arabic LMM benchmark. Findings of NAACL 2025. arXiv:2410.18976.

Guan, T., Liu, F., Wu, X., Xian, R., Li, Z., Liu, X., Wang, X., Chen, L., Huang, F., Yacoob, Y., Manocha, D., & Zhou, T. (2024). HallusionBench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. CVPR 2024. arXiv:2310.14566.

Li, Y., Du, Y., Zhou, K., Wang, J., Zhao, W. X., & Wen, J.-R. (2023). Evaluating object hallucination in large vision-language models. EMNLP 2023. arXiv:2305.10355.

Lin, S., Hilton, J., & Evans, O. (2022). TruthfulQA: Measuring how models mimic human falsehoods. ACL 2022. arXiv:2109.07958.

Myung, J., Lee, N., Zhou, Y., Jin, J., Putri, R. A., Antypas, D., Borkakoty, H., Kim, E., Perez-Almendros, C., Ayele, A. A., Gutiérrez-Basulto, V., Ibáñez-García, Y., Lee, H., Muhammad, S. H., Park, K., Rzayev, A. S., White, N., Yimam, S. M., Pilehvar, M. T., Ousidhoum, N., Camacho-Collados, J., & Oh, A. (2024). BLEnD: A benchmark for LLMs on everyday knowledge in diverse cultures and languages. NeurIPS 2024 Datasets & Benchmarks Track. arXiv:2406.09948.

Rao, A., Yerukola, A., Shah, V., Reinecke, K., & Sap, M. (2025). NormAd: A framework for measuring the cultural adaptability of large language models. NAACL 2025. arXiv:2404.12464.

Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E. P., Zhang, H., Gonzalez, J. E., & Stoica, I. (2023). Judging LLM-as-a-judge with MT-Bench and Chatbot Arena. NeurIPS 2023 Datasets and Benchmarks Track. arXiv:2306.05685.

Note on Pearl and PalmX 2025: both share a first author (Alwajih) and a publication year (2025); they're disambiguated in-text as "(Alwajih, Magdy, et al., 2025)" for Pearl and "(Alwajih, El Mekki, et al., 2025)" for PalmX, rather than both collapsing to the same ambiguous "(Alwajih et al., 2025)."

A note on AI assistance. I had help from Claude Code with the coding in this project.