ArabicNLPWorld/arabic-islamic-hallucination-synthetic

🧪 IslamicEval 2026 Synthetic Subtask 2 Dataset

Description

Synthetic training dataset for IslamicEval 2026 Subtask 2 (Hallucination Identification).
Created from the Canonical Islamic Corpus by applying realistic Arabic/Islamic distortions.

• Correct examples: original canonical texts.
• Incorrect examples: texts with one or more realistic errors (word replacements, swaps, deletions, character changes, negation removal, number changes, etc.).

📊 Dataset Statistics

Metric	Value
Total entries	543,257
Correct	322,149
Incorrect	221,108
Quran fragments	10,539
Hadith fragments	532,718

🧩 Applied Distortion Types

• replace_word_realistic (98,184)
• delete_word (48,612)
• insert_word_realistic (36,616)
• swap_parts_hadith (27,440)
• delete_negation (4,878)
• swap_adjacent_words (3,869)
• change_number (775)
• char_change (734)

📦 Data Fields

Field	Type	Description
id	string	Unique example ID (e.g., 123_correct, 123_dist_1)
original_id	int32	ID of the source canonical text
text_original	string	Original authentic text
text	string	Text to classify (canonical or distorted)
label	string	correct or incorrect
distortion_types	string	JSON list of applied distortions (e.g., ["replace_word_realistic"])
source	string	Origin dataset (tanzil, fawazahmed0/hadith-data, etc.)
type	string	quran or hadith
surah	int32	Surah number (Quran only, else 0)
ayah	int32	Ayah number (Quran only, else 0)
collection	string	Hadith collection name
section	string	Chapter/section
hadith_number	string	Hadith number

📥 Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("ArabicNLPWorld/islamiceval2026-synthetic-subtask2", split="train")

# Correct example
for ex in dataset.filter(lambda x: x['label'] == 'correct').select(range(2)):
    print(f"[CORRECT] {ex['text']}")

# Incorrect example with distortion types
for ex in dataset.filter(lambda x: x['label'] == 'incorrect').select(range(2)):
    print(f"[INCORRECT] {ex['text']}")
    print(f"Distortions: {ex['distortion_types']}")

📜 Source

Generated from Canonical Islamic Corpus.
License: CC BY-SA 4.0

🏷️ Citation

@misc{mullosharaf_synthetic_subtask2_2026,
  author = {Mullosharaf Arabov},
  title = {IslamicEval 2026 Synthetic Subtask 2 Dataset},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/ArabicNLPWorld/islamiceval2026-synthetic-subtask2}}
}

---

Prepared with dedication for IslamicEval 2026. Inshallah, victory!