IFM/PalmX2025-Subtask2_verified

PalmX 2025, Subtask 2: Islamic culture (verified)

A filtered version of the PalmX 2025 Islamic Culture evaluation set (multiple-choice questions on Islamic knowledge), with samples flagged by a multi-judge audit removed.

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 PalmX questions. The source dataset is gated and licensed for non-commercial research only, so we publish the filter instead: the list of source ids to keep, plus a script that applies it to your own copy of the source.

How to use

1. Request access to UBC-NLP/palmx_2025_subtask2_islamic and accept its terms.
2. Log in with hf auth login.
3. Run apply_filter.py, which writes the 948-row verified test set:

pip install datasets pandas huggingface_hub
python apply_filter.py palmx2025_subtask2_verified.parquet

Or filter it yourself:

import pandas as pd
from datasets import load_dataset

ids = pd.read_csv("hf://datasets/IFM/PalmX2025-Subtask2_verified/ids.csv")
kept = set(ids.loc[ids["status"] == "kept", "id"])
test = load_dataset("UBC-NLP/palmx_2025_subtask2_islamic", split="test",
                    revision="95864e9a18753a2aab5b3b687a494a8638751606")
verified = test.filter(lambda row: row["id"] in kept)  # 948 rows

Files

• ids.csv: one row per source test id (1,001 rows) with a status column:
  • kept (948): part of the verified set.
  • removed (37): excluded by the audit. The judges' labels, reasons and suggested fixes are in IFM/discarded_samples, config PalmX 2025-2, keyed by the same id.
  • no_gold_upstream (16): the source row has an empty answer, so it was never part of the evaluation set.
• apply_filter.py: rebuilds the verified set from the source.

Source

• Original dataset: UBC-NLP/palmx_2025_subtask2_islamic (gated), revision 95864e9a18753a2aab5b3b687a494a8638751606, test split (1,001 rows). Website: https://palmx.dlnlp.ai/.
• The evaluation set is the 985 test rows that have a gold answer.
• Two question texts appear twice upstream with different ids and options (ids 69/517 and 195/784); both copies are kept unless the audit excluded them.

Audit and filtering

Samples were screened with the multi-judge audit described in IFM/discarded_samples (revision 3a3c1339, config PalmX 2025-2). 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). Rows that were not excluded are unchanged and keep their original order.

Split	Evaluation rows	Excluded	Kept
test	985	37	948

License and access

The source dataset is distributed under the PalmX Data License (non-commercial research use), and access requires accepting its terms on the Hub. This repository redistributes no source text, only source ids and audit decisions; the verified set you build with it remains subject to the PalmX license.

Citation

Please cite the shared-task paper:

@misc{alwajih2025palmx2025sharedtask,
  title         = {PalmX 2025: The First Shared Task on Benchmarking LLMs on Arabic and Islamic Culture},
  author        = {Fakhraddin Alwajih and Abdellah El Mekki and Hamdy Mubarak and Majd Hawasly and Abubakr Mohamed and Muhammad Abdul-Mageed},
  year          = {2025},
  eprint        = {2509.02550},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2509.02550}
}

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