ahmad1979/ahmad-eval-v1

Ahmad-Eval v1

A 232-question retrieval and QA benchmark for Quranic question answering. Queries are in Modern Standard Arabic, English, and Romanized Urdu. Gold-standard verse references are always to the Arabic Quran (6236 verses, Uthmani script).

This dataset is intended as an evaluation benchmark for retrieval and question-answering systems built on the Quran. It is not a fiqh (Islamic jurisprudence) dataset, and it is not intended to answer religious questions authoritatively. See "Intended uses" and "Out-of-scope uses" below.

Dataset description

Ahmad-Eval v1 contains 232 questions across six evaluation buckets, each testing a distinct retrieval and abstention capability. The questions are hand-labeled or drawn from public sources; the gold verse sets are human-verified. Each question carries a source_dataset field identifying its origin.

The dataset was built in Sessions 6–7 of an 8-session development cycle to evaluate a retrieval + verifier + scope-routing pipeline on the Quran. The pipeline's measured results on this benchmark, including its observed failure modes, are in pipeline_results.md. They are linked from this card so that anyone using ahmad_eval_v1 as a baseline sees both the benchmark and the pipeline that produced it, with its limitations recorded honestly.

Composition

Total: 232 questions.

By bucket:

bucket	n	lang	tests
lexical_ar	66	ar	root or surface keyword matching
cross_lingual_en	46	en	cross-lingual retrieval from English to Arabic
roman_urdu	45	ur-roman	low-resource transliteration retrieval
semantic_ar	45	ar	paraphrase and thematic matching
unanswerable_ar	22	ar	correct abstention (out-of-corpus questions)
multi_hop_ar	8	ar	compound queries combining two concepts

By source:

source_dataset	n	provenance
manual	196	hand-labeled in Sessions 6–7 of this project
synthetic_gemini_v2	36	contributed by an external model, hand-reviewed, frozen before evaluation (Session 7k)

The 36 synthetic_gemini_v2 questions are the "fresh" evaluation set. They were not present when the pipeline's retrieval stack was tuned. The 196 manual questions are the "calibration" set.

Fields

Each question in ahmad_eval_v1.json is an object with these fields:

field	type	description
qid	string	unique question identifier
lang	string	query language: ar, en, or ur-roman
bucket	string	one of the six bucket names above
query	string	question text in the stated language
expected_verses	array of strings	gold-standard verse refs in "S:A" format; empty for unanswerable_ar
answerable	boolean	false iff bucket == "unanswerable_ar"
source_dataset	string	"manual" or "synthetic_gemini_v2"
source_id	string	original ID within source, "" for manual
notes	string	labeling notes and adversarial properties

Splits

• test split: all 232 questions. There is no train split. This is an evaluation benchmark; using it for training contaminates it as a fresh-question test.

To reproduce the "calibration" / "fresh" split used in the reference pipeline's evaluation, filter on source_dataset: manual for calibration (196 questions), synthetic_gemini_v2 for fresh (36 questions).

Intended uses

• Evaluating retrieval systems built on the Quran. R@5, R@10, MRR, hit@k, and similar metrics on the retrieval stack.
• Evaluating abstention behavior on out-of-corpus questions. The unanswerable_ar bucket is designed to test whether a system correctly refuses to answer questions whose rulings are not in the Quran.
• Evaluating cross-lingual retrieval from English and Romanized Urdu to Arabic verses.
• Reproducing the reference pipeline's evaluation, using the artifacts described in pipeline_results.md.

Out-of-scope uses

• Answering religious questions authoritatively. This is an evaluation benchmark, not a fatwa service. The gold verses are retrieval targets, not rulings. Anyone building a user-facing Islamic QA system on this data must route rulings to qualified scholars, as the reference pipeline's gate does.
• Training a model to answer Quranic questions. The 232 questions are evaluation data. Training on them contaminates every future evaluation on this benchmark.
• Publishing a "system accuracy" number from this benchmark without also publishing the abstention split and the failure-mode analysis. A verified rate without the abstention rate is not informative; see pipeline_results.md for why.
• Claiming the gold expected_verses are the only valid answers. They are a set of valid verses, not the set. The benchmark's gold sets are known to be narrower than the space of relevant verses; see "Known limitations" below.

Known limitations and observed failure modes

These are recorded here so that anyone using ahmad_eval_v1 as a baseline understands what the reference pipeline found when it ran on this benchmark, and what a better system would need to do differently.

Limitation 1 — the gold sets are narrow

The expected_verses for each question are a specific set of verses that answer the question. They are not the only such verses. A retrieval system that returns a valid verse not in expected_verses will be scored as "wrong" by strict R@5, even though the verse is a valid answer.

This was quantified in Day 28 of the project's development (see journal/2026-10-18-day28.md): a hand-review of 15 sampled retrieval failures found that 13 of 15 (87%) returned at least one verse a human reviewer marked as clearly or plausibly relevant. The failures were gold-set narrowness, not retriever failures, in the majority of cases.

Implication for evaluation: R@5 on this benchmark understates a system's real retrieval quality. If evaluating a new system, report R@5 and hit@5 and (ideally) a hand-reviewed "relevance@5" sample. Do not report R@5 alone as a retrieval- quality number.

Limitation 2 — the unanswerable_ar bucket is out-of-corpus, not unanswerable in principle

The 22 unanswerable_ar questions include two subtypes:

• type (a): questions the Quran does not address at all (e.g. "who won the 1998 World Cup").
• type (b): questions whose topic appears in the Quran but whose ruling comes from hadith or fiqh (e.g. the number of rak'ahs in witr prayer, the ruling on smoking).

Type (b) questions are not unanswerable in Islamic scholarship; they are out-of-corpus for a Quran-only retrieval system. A system that answers type (b) from a broader corpus would be correct in principle. The bucket name is unanswerable_ar for historical reasons; out_of_corpus would be more accurate.

Implication for evaluation: treat the unanswerable_ar bucket as "should abstain given a Quran-only corpus," not as "should abstain in all contexts."

Limitation 3 — the query phrasing distribution is not uniform

The 196 manual questions and the 36 synthetic_gemini_v2 questions are phrased differently. The manual set contains a large proportion of single-word Arabic lexical queries (التوبة, الصبر). The synthetic_gemini_v2 set contains proportionally more full-sentence prose questions, in all three languages.

This is not a limitation of the benchmark per se — diversity of phrasing is a feature. But it means that a system tuned on the 196 will not necessarily generalize to the 36, and this was observed in the reference pipeline: a Layer 1 classification prompt that over-fired on 18% of the 196's answerable questions over-fired on 57% of the 36 fresh questions. See pipeline_results.md for the full analysis.

Implication for evaluation: if your system was developed with the 196 visible, evaluate it on the 36 as a fresh test, not on the 196 as a confirmation set.

Limitation 4 — cross-lingual translations are not perfect

The cross_lingual_en and roman_urdu questions depend on machine translation or transliteration mappings for their Arabic query form. Some translations are non-literal (e.g. "backbiting" in the Yusuf Ali English corpus is rendered "behind their backs" in Surah 49), and a system that relies on literal translation for cross-lingual retrieval will fail on such cases.

This is documented in the Day 28 journal as a specific stemmer-vocabulary mismatch on the Roman-Urdu ghibat kya hai question.

Observed failure mode — the reference pipeline

The reference pipeline (retrieval + verifier + two-layer gate) was run on this benchmark in Sessions 7h–7k. The measured results are in pipeline_results.md. The short version:

• On the 196 calibration questions, the pipeline achieved 32 verified, 33 abstained, 115 wrong on the 180 answerable; 16/16 out-of-scope abstained, 0 false answers.
• On the 36 fresh questions, the pipeline achieved 3 verified, 17 abstained, 10 wrong on the 30 answerable; 4/6 out-of-scope abstained, 2 false answers.

The 36-fresh numbers are the publishable result. They fail the pre-registered thresholds the reference pipeline committed to before running. The dominant failure is that the reference pipeline's scope-routing gate over-fires on the fresh set's prose-like phrasing: 17 of 30 answerable fresh questions were gated out before reaching the retriever. See pipeline_results.md for the analysis.

Implication for evaluation: the reference pipeline is a useful baseline for retrieval on the 196, and an honest failure case for generalization on the 36. Any new system that improves on the reference pipeline's 36-fresh numbers has genuinely improved generalization.

Licensing and provenance

• The Quranic text (canonical_quran.db in the reference pipeline) is public domain.
• The English translation is the Yusuf Ali translation, public domain.
• The 196 manual questions were hand-labeled in this project.
• The 36 synthetic_gemini_v2 questions were contributed by an external model and hand-reviewed before being frozen in the benchmark. They are not produced by the reference pipeline's authors; they were reviewed for correctness against the Uthmani text.
• The dataset is released under CC-BY-4.0.

Citation

If you use ahmad_eval_v1 in published work, cite it as: Ahmad-Eval v1: A 232-question retrieval and QA benchmark for the Quran. Project: Ahmad-AI. Sessions 6–7, 2026. Repository: https://github.com/emadaed/quran-retrieval-benchmark

See also

• pipeline_results.md — full pipeline evaluation results, failure-mode analysis, confidence intervals, and pre-registered pass-rule verdict. (In this dataset repo.)
• Journal: Day 28, Session 7h — the template-drafter agent (first end-to-end pipeline). (In the code repo.)
• Journal: Day 28, Session 7g — development notes on the retriever's measurements and the gold-set-narrowness finding. (In the code repo.)
• Journal: Day 30, Session 7j — the reference pipeline's calibration run (196 questions). (In the code repo.)
• Journal: Day 31, Session 7k — the reference pipeline's fresh-question run (36 questions) and the failure analysis. (In the code repo.)