IFM/QIAS2025_verified

QIAS 2025: Islamic inheritance and general Islamic knowledge MCQs (verified)

A filtered version of the multiple-choice data of the QIAS 2025 shared task (Islamic inheritance reasoning and knowledge assessment), with samples flagged by a multi-judge audit removed. Two tasks: task1 (Task 1, Islamic inheritance: test set with gold labels) and task2 (Task 2, general Islamic knowledge: development set).

What "verified" means. Samples were removed when all three LLM judges flagged a significant defect. This is not a human review, and only samples that baseline models answered incorrectly were audited, so some defects may remain in the kept data.

This repository does not contain the QIAS questions. The source data is public but states no license, so we publish the filter instead: which source rows to keep, plus a script that applies it to the source.

How to use

pip install pandas pyarrow huggingface_hub
python apply_filter.py qias_verified

apply_filter.py downloads both source files at the commit below, checks every row against the fingerprints in ids.csv, and writes the verified sets task1.parquet (898 rows) and task2.parquet (689 rows). It stops with an error if the source has changed.

Or filter it yourself:

import pandas as pd

BASE = "https://gitlab.com/islamgpt1/qias_shared_task_2025/-/raw/98e6cc179ac5afeff88f8d219a084a79d117b0ee/data"
FILES = {"task1": "Task1/Task1_MCQ_Test_gold_labels.csv", "task2": "Task2/Task2_MCQs_Dev.csv"}

ids = pd.read_csv("hf://datasets/IFM/QIAS2025_verified/ids.csv")
verified = {}
for task, path in FILES.items():
    source = pd.read_csv(f"{BASE}/{path}", encoding="utf-8-sig", dtype=str, keep_default_na=False)
    keep = ids[ids["task"].eq(task)]["status"].eq("kept").to_numpy()
    verified[task] = source[keep].reset_index(drop=True)  # task1: 898 rows, task2: 689 rows

Files

• ids.csv: one row per source row (1,700 rows):
  • task: task1 or task2.
  • row: the 0-based row index in that task's source file.
  • id_question: the source id_question.
  • status: kept (1,587) or removed (113).
  • fingerprint: the first 16 hex characters of a SHA-256 over all of the row's cells (read as text, joined by \x1f), used to detect changes in the source.
• apply_filter.py: rebuilds the verified sets from the source.

Source

• Original data: GitLab islamgpt1/qias_shared_task_2025, commit 98e6cc179ac5afeff88f8d219a084a79d117b0ee: data/Task1/Task1_MCQ_Test_gold_labels.csv (1,000 rows) and data/Task2/Task2_MCQs_Dev.csv (700 rows).
• Columns of the rebuilt sets, as in the source: id_question, question, option1 ... option6 (task1) or option1 ... option4 (task2), label (gold answer) and level (difficulty).
• The rebuilt sets keep the source text exactly as published, including its whitespace. An earlier evaluation copy of this dataset differed from the source in whitespace only, in 34 rows: 20 task1 rows whose option cells contain carriage returns in the source, and 14 task2 rows with non-breaking spaces in the source where the copy had ordinary spaces. No cell differs once whitespace is normalised.

Audit and filtering

Samples were screened with the multi-judge audit described in IFM/discarded_samples (configs QIAS-1 for task1 and QIAS-2 for task2). Three language-model judges (GPT-5.6 Sol, Gemini 3.8 Flash, Claude Opus 5) assessed candidate samples that strong baselines answered incorrectly. A sample is excluded only when all three judges report a significant issue (a problem with the question, the options or the gold label that prevents reliable answering or scoring). Each discarded record's doc_id is the 0-based row index in its task's source file; ids.csv gives the matching source row (row) for each excluded sample. Every match was checked against the record's question text, options, gold label and difficulty level. All other rows are kept unchanged and in their original order.

Task	Level	Original rows	Excluded	Kept
task1	Beginner	500	98	402
task1	Advanced	500	4	496
task1	All	1,000	102	898
task2	beginner	350	6	344
task2	intermediate	175	3	172
task2	advanced	175	2	173
task2	All	700	11	689
Total		1,700	113	1,587

License

The source repository states no license for the data, so this repository redistributes none of its text. The files here (ids.csv, apply_filter.py and this card) are released under CC BY 4.0. The verified sets you build from the source remain subject to the source's terms.

Citation

Please cite the shared-task overview (BibTeX as given in the source repository):

@inproceedings{bouchekif2025qias,
  title     = {QIAS 2025: Overview of the Shared Task on Islamic Inheritance Reasoning and Knowledge Assessment},
  author    = {Bouchekif, Abdessalam and Rashwani, Samer and Mohamed, Emad and Al-Khatib, Mutaz and Sbahi, Heba and Gaben, Shahd and Zaghouani, Wajdi and Erbad, Aiman and Ghaly, Mohammed},
  booktitle = {Proceedings of The Third Arabic Natural Language Processing Conference (ArabicNLP 2025)},
  address   = {Suzhou, China},
  month     = {November},
  year      = {2025},
  publisher = {Association for Computational Linguistics}
}

@misc{jais3_2026,
  title  = {Jais 3: Introducing Arabic-centric Agentic AI},
  author = {{IFM Team}},
  year   = {2026},
  url    = {https://ifm.ai/blog/jais3/},
}