QCRI/Fanar-Sadiq-Bench

Fanar-Sadiq Classifier Datasets

Fanar-Sadiq Classifier Datasets is a bilingual Arabic/English collection of query-classification datasets released with the paper “Fanar-Sadiq: A Multi-Agent Architecture for Grounded Islamic QA.”

The datasets support the routing components of Fanar-Sadiq, a multi-agent Islamic question-answering system that routes user queries to specialized modules such as retrieval-grounded Islamic QA, Quran retrieval, du'a lookup, prayer-time computation, Islamic calendar tools, zakat calculators, and inheritance calculators.

This repository contains two classifier datasets:

1. Prompt classifier: a binary dataset for deciding whether a user query is Islamic or non-Islamic.
2. Intent classifier: a fine-grained Islamic intent dataset for routing Islamic queries to the appropriate specialized component.

Dataset Summary

Subset	Rows	Task	Labels
prompt classifier	~1.72k	Binary Islamic query detection	islamic, non_islamic
intent classifier	700	Fine-grained Islamic intent classification	9 Islamic intent labels

The data includes both Arabic and English queries, including naturally phrased, code-switched, short, long, factual, jurisprudential, conversational, and tool-oriented user requests.

Intended Use

This dataset is intended for research and development on:

• Islamic query classification
• Arabic/English text classification
• Query routing for multi-agent or tool-augmented QA systems
• Intent classification for religious-domain assistants
• Evaluation of bilingual Islamic QA routing pipelines

Typical use cases include training or evaluating models that decide:

• whether an incoming query should be handled by an Islamic QA system;
• which specialized Islamic QA module should answer the query;
• whether a query requires retrieval, deterministic computation, or structured tool use.

Dataset Subsets

1. Prompt Classifier

The prompt classifier subset is a binary classification dataset used to determine whether a query is related to Islamic content.

Schema

Field	Type	Description
id	string	Unique example identifier
query	string	User query in Arabic or English
label	string	Binary label: islamic or non_islamic

Labels

Label	Description
islamic	Query is related to Islam, Islamic knowledge, Quran, Hadith, fiqh, Islamic history, worship, religious practice, or Islamic-domain reasoning
non_islamic	Query is general-purpose, unrelated to Islamic QA, or should be routed to a general assistant

Example

{
  "id": "268fe685",
  "query": "قارن بين تنظيم الحياة العامة وفق الشريعة الإسلامية والثقافة الغربية .",
  "label": "islamic"
}

{
  "id": "9c5eae33",
  "query": "لدي بنطالاً أبيض، ما لون القميص الذي يليق عليه؟",
  "label": "non_islamic"
}

2. Intent Classifier

The intent classifier subset is a fine-grained Islamic intent classification dataset used to route Islamic queries to specialized components in Fanar-Sadiq.

The dataset was created from real user queries, anonymized and filtered to remove personally identifying information. Each query was annotated by three independent annotators, and the final label was determined by majority vote.

Schema

Field	Type	Description
id	string	Unique example identifier
query	string	User query in Arabic or English
label	string	Final intent label selected by majority vote

Intent Labels

Label	Description
fiqh_ruling	Questions about Islamic legal rulings, permissibility, obligations, halal/haram, or jurisprudence
quran_retrieval	Requests for Quranic verses, ayahs, surahs, Quran text, or Quran-related retrieval
general_islamic	General Islamic knowledge, history, concepts, beliefs, personalities, events, or explanations
greeting	Greetings, pleasantries, and short conversational openings
zakat_calculation	Questions requiring zakat computation or zakat-related financial reasoning
inheritance_calculation	Islamic inheritance/faraid questions requiring structured calculation
dua_lookup	Requests for supplications, adhkar, or du'a for specific occasions
islamic_calendar	Hijri/Gregorian conversion, Islamic dates, Ramadan/Eid dates, or Islamic events
prayer_times	Prayer times, Qibla direction, sunrise/sunset, or location/time-based worship queries

Example

{
  "id": "2772",
  "query": "Jazakallah khair for the help!",
  "label": "greeting",
}

{
  "id": "83123",
  "query": "ما هو نصيب الأخت الشقيقة مع وجود ابنتين للمتوفى؟ هل تعتبر عصبة؟",
  "label": "inheritance_calculation",
}

Loading the Dataset

You can load the dataset with the Hugging Face datasets library.

from datasets import load_dataset

# List available configurations
# from datasets import get_dataset_config_names
# print(get_dataset_config_names("QCRI/Fanar-Sadiq-Classifier-Datasets"))

intent_ds = load_dataset(
    "QCRI/Fanar-Sadiq-Classifier-Datasets",
    name="intent classifier"
)

prompt_ds = load_dataset(
    "QCRI/Fanar-Sadiq-Classifier-Datasets",
    name="prompt classifier"
)

print(intent_ds)
print(prompt_ds)

Each subset currently contains a test split.

print(intent_ds["test"][0])
print(prompt_ds["test"][0])

Baseline Task Formulations

Binary Classification

For the prompt classifier subset:

Input: query
Output: islamic | non_islamic

Example:

Query: ما حكم الزكاة في الإسلام؟
Label: islamic

Multi-Class Intent Classification

For the intent classifier subset:

Input: query
Output: one of the 9 Islamic intent labels

Example:

Query: متى أذان الفجر في مكة المكرمة غداً؟
Label: prayer_times

Relationship to Fanar-Sadiq

Fanar-Sadiq is a bilingual Arabic/English multi-agent Islamic QA architecture. Instead of sending every query through a single retrieve-then-generate pipeline, Fanar-Sadiq first classifies the query and routes it to specialized components.

Examples of routes include:

• retrieval-grounded fiqh and Islamic knowledge QA;
• exact Quran verse lookup and Quran-related routing;
• deterministic zakat and inheritance calculators;
• prayer-time and Qibla tools;
• Islamic calendar conversion and event lookup;
• du'a lookup and greeting handling.

These datasets are designed to evaluate and support the routing layer of that architecture.

Annotation

Intent Classifier

The intent-classification dataset was developed from anonymized real user queries. Each query was labeled by three independent annotators into one of nine intent categories. The final label was selected by majority vote, and examples without majority agreement were discarded.

The released intent subset includes 700 examples with majority labels.

Prompt Classifier

The binary prompt-classifier dataset supports the orchestrator that decides whether a query should be routed to Fanar-Sadiq or to a general-purpose assistant. Queries labeled as Islamic are routed to the Islamic QA system; other queries are handled by a general assistant.

Ethical Considerations and Limitations

This dataset is intended for classifier training and evaluation, not for issuing religious rulings.

Important limitations:

• The dataset contains user-style queries and may include sensitive religious, legal, social, or ethical topics.
• Labels reflect the routing schema used by Fanar-Sadiq and should not be interpreted as religious judgments.
• A query labeled fiqh_ruling means that the query asks about jurisprudence; it does not imply any answer or ruling.
• Models trained on this dataset may make routing errors, especially for ambiguous, multi-intent, or code-switched queries.
• Islamic legal questions may involve scholarly disagreement, madhhab-specific assumptions, local practice, and contextual details not present in the query.
• This dataset should not be used as a replacement for qualified scholarly expertise.

Citation

If you use this dataset, please cite the Fanar-Sadiq paper:

@inproceedings{abbas2026fanarsadiq,
  title     = {{Fanar-Sadiq}: A Multi-Agent Architecture for Grounded Islamic QA},
  author    = {Abbas, Ummar and Ouzzani, Mourad and Eltabakh, Mohamed Y. and Sinan, Omar and Bhatia, Gagan and Mubarak, Hamdy and Hawasly, Majd and Hashim, Mohammed Qusay and Darwish, Kareem and Alam, Firoj},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics},
  month     = jul,
  year      = {2026},
  address   = {San Diego, California, United States},
  publisher = {Association for Computational Linguistics},
  note      = {Accepted to ACL 2026}
}

License

This dataset is released under the Apache License 2.0.