imenselmi/Elm-NLP-Challenge-Mushroom-Dataset

Elm NLP Challenge Mushroom Dataset A comprehensive Arabic hallucination detection dataset designed for evaluating and improving Large Language Models' factual accuracy across diverse Arabic domains.

📊 Dataset Overview Dataset Size: 10,000 samples Language: Arabic Task: Hallucination Detection & Evaluation Format: JSONL 🎯 Purpose This dataset is specifically designed to test and evaluate hallucination in Arabic language models across multiple domains, providing a robust benchmark for assessing factual accuracy and reliability of LLM outputs in Arabic contexts.

📁 Dataset Structure Each entry in the dataset contains the following keys: { "id": "unique_id_020483", "prompt": "اكتب سيرة ذاتية لريك ماهلر", "reference_answer": "ريتشارد كيث ماهلر..." } keys Description: id: Unique identifier for each sample prompt: The input question/prompt in Arabic reference_answer: The ground truth answer for hallucination detection

🔍 Domain Coverage The dataset encompasses a rich diversity of Arabic content across the following domains: general Quranic Questions Arabic Poetry: Complex poetic forms with challenging meters including: خفيف (Khafif) مديد (Madid) منسرح (Munsarih) مجتث (Mujtath) Arabic Grammar: Linguistic and grammatical questions Legal Questions: Legal and law-related queries Educational Content: School-level questions across various fields Biography: Historical and contemporary biographical information Geographic Coverage: Questions about Arabic countries including: Sudan Saudi Arabia Morocco

🖥️ Hardware Specifications GPU Configuration GPU Model: NVIDIA H100 80GB HBM3 Number of GPUs: 8 Memory per GPU: 81,559 MiB (~80 GB) Total GPU Memory: ~640 GB

Software Environment : NVIDIA Driver Version: 535.183.01 CUDA Version: 12.2 Operating System: Ubuntu

🤖 Models Used Primary Model Model: Qwen/Qwen3-14B Role: Primary model for testing and evaluation Judge Model Model: google/gemma-3-27b-it Role: Judge model for hallucination evaluation

📏 Evaluation Methodology method 1: Primary Metric: Cosine Similarity Threshold: 85% The evaluation process involves: Pass the questions to Qwen-3 14B, embed them using SentenceTransformer("all-MiniLM-L6-v2"), and then use cosine similarity with an 0.85 threshold to evaluate the Qwen output against the ground truth.So Computing cosine similarity scores. method 2: Evaluation Approach: Comparative analysis using Gemma as judge model The evaluation process involves: Generating responses using Qwen3-14B Comparing generated outputs against reference answers Using Gemma-3-27B-IT as an additional judge for quality assessment

📚 References & Acknowledgments This dataset builds upon and draws inspiration from the following resources:

Wiki Bio GPT3 Hallucination Dataset MetRec: A dataset for meter classification of arabic poetry QIAS 2025 Islamic Assessment Studies Humanity’s Last Exam NLP Project LLM Hallucination MizanQA-v0 ALRAGE Dataset PALMX 2025 Culture Subtask

🚀 Usage python from datasets import load_dataset

Load the dataset

dataset = load_dataset("imenselmi/Elm-NLP-Challenge-Mushroom-Dataset")

Access a sample

sample = dataset['train'][0] print(f"Prompt: {sample['prompt']}") print(f"Reference: {sample['reference_answer']}")

📄 License [Please specify your license here]

🙏 Citation If you use this dataset in your research, please cite:

bibtex @dataset{elm_nlp_challenge_mushroom, title={Elm NLP Challenge Mushroom: An Arabic Hallucination Detection Dataset}, author={Imen Selmi, Alaa Dania Asimi, Assala BenMalik}, year={2026}, publisher={Hugging Face}, howpublished={https://huggingface.co/imenselmi/Elm-NLP-Challenge-Mushroom-Dataset} } Dataset Version: 1.0 Last Updated: January 2026