Hallucination Detection Dataset / Hallucinated Entity Recognition (HER) Dataset

A comprehensive, cross-domain dataset with span-level annotations for training fine-grained hallucination detection models

🎯 Overview

As Large Language Models (LLMs) become increasingly prevalent, hallucinations in their outputs pose serious threats to society by propagating misinformation. This repository contains the first cross-domain, balanced dataset specifically designed for training span-level hallucination detection models.

Key Features

• 22,912 labeled samples across 6,028 topics
• 8 diverse domains: Politics, Science, Healthcare, Law, History, Geography, Sports, Media
• Span-level annotations with supporting evidence
• Challenging hallucinations (only 17% human detection accuracy)
• Balanced dataset with both true and false claims
• Domain-agnostic generation pipeline

📊 Dataset Statistics

Metric	Value
Total Claims	22,912
True Claims	13,031 (56.9%)
False Claims	9,881 (43.1%)
Topics Covered	6,028
Domains	8
Human Detection Accuracy	17%

🏗️ Dataset Structure

Fields Description

{
    "context": str,      # Text that needs hallucination checking
    "evidence": str,     # Supporting/refuting evidence
    "hf": bool,         # Hallucination flag (1=hallucinated, 0=true)
    "hinfo": {          # Detailed hallucination information
        "hp": str,       # Hallucinated phrases
        "hs": [int, int], # Character spans of hallucinations
        "pp": [str, str], # Original→replacement phrase mappings
        "cp": [str, str], # Context→paraphrase sentence mappings  
        "sf": [str]      # Sample factual phrases
    }
}

Example Entry

{
    "context": "Following these withdrawals the Bid Central Committee met on 9 June 2017 in Lausanne Switzerland to discuss the 2024 and 2028 bid processes",
    "evidence": "Following these withdrawals the IOC Executive Board met on 9 June 2017 in Lausanne Switzerland to discuss the 2024 and 2028 bid processes", 
    "hf": 1,
    "hinfo": {
        "hp": "Bid Central Committee",
        "hs": [28, 49],
        "pp": ["IOC Executive Board", "Bid Central Committee"],
        "cp": ["original sentence", "paraphrased evidence"],
        "sf": ["2024 and 2028 bid processes"]
    }
}

🔬 Methodology

Our synthetic data generation pipeline addresses key limitations in existing datasets:

Pipeline Components

1. Claim Extraction - Identifies verifiable claims using YAKE-based classification
2. Key-phrase Extraction - Extracts objective, replaceable phrases
3. Subjective Filtering - Removes opinion-based content using POS tagging + SentiWordNet
4. Semantic Replacement - Generates plausible alternatives via masked language modeling
5. Entailment Verification - Ensures semantic contradiction using RoBERTa-MNLI
6. Evidence Generation - Creates supporting evidence through controlled paraphrasing

Quality Assurance

• Textual Entailment Check: Ensures replacements create genuine contradictions
• Semantic Reranking: Maximizes semantic distance while preserving plausibility
• Human Evaluation: 17% detection accuracy confirms challenging, realistic hallucinations

📈 Evaluation Results

Human Study Results

• Overall Accuracy: 17% (challenging for humans)
• Hallucination Detection: 26% when hallucination present
• True Negative Rate: 8% when no hallucination present

Evidence Quality Distribution

Quality Level	Count	Percentage	Jaccard Score Range
High	451	34.1%	≥ 0.83
Medium	782	57.3%	0.55 - 0.83
Low	131	9.6%	< 0.55

🚀 Getting Started

Installation

pip install datasets transformers torch

Loading the Dataset

from datasets import load_dataset

# Load from Hugging Face
dataset = load_dataset("your-username/hallucination-detection-dataset")

# Access training data
train_data = dataset['train']
test_data = dataset['test']

print(f"Training samples: {len(train_data)}")
print(f"Test samples: {len(test_data)}")

Basic Usage Example

import json

# Load a sample
sample = train_data[0]

print("Context:", sample['context'])
print("Evidence:", sample['evidence']) 
print("Contains Hallucination:", bool(sample['hf']))

if sample['hf']:
    hinfo = json.loads(sample['hinfo'])
    print("Hallucinated Phrase:", hinfo['hp'])
    print("Character Span:", hinfo['hs'])

🧪 Baseline Models

We provide baseline implementations for:

• Span-level Detection: Identifies hallucinated text segments
• Binary Classification: Determines presence of hallucinations
• Evidence-based Verification: Leverages supporting evidence

Training Example

from transformers import AutoTokenizer, AutoModelForTokenClassification

# Load model for span-level detection
model_name = "roberta-base" 
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name, num_labels=3)

# BIO tagging: B-HAL, I-HAL, O
# Your training code here...

📚 Domain Coverage

Domain	Topics	Description
Politics	~750	Government, elections, policies
Science	~750	Research papers, discoveries, theories
Healthcare	~750	Medical conditions, treatments, research
Law	~750	Legal cases, regulations, procedures
History	~750	Historical events, figures, dates
Geography	~750	Countries, landmarks, demographics
Sports	~750	Athletes, competitions, records
Media	~750	News, entertainment, broadcasting

🔄 Mathematical Formalism

The paper introduces a formal framework for efficient hallucination detection:

Problem Formulation

• Claims Decomposition: Text T → Claims C = {c₁, c₂, ..., cₙ}
• Atomic Claims: Each claim cᵢ → Atomic claims {Aᵢⱼ}
• Evidence Retrieval: Optimize weighted coverage with budget constraints
• Iterative Verification: Update verification status until convergence

Cost-Sensitive Detection

E* = arg max Σ w(Aᵢⱼ)
     E'⊆E    Aᵢⱼ∈⋃ₑₖ∈E' C(eₖ)
     
subject to |E'| ≤ k

📖 Citation

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

@article{kumar2025synthetic,
  title={Generating Synthetic Data for Hallucination Detection in LLMs},
  author={Kumar, Rohit and Chatterjee, Niladri},
  journal={arXiv preprint arXiv:XXXX.XXXXX},
  year={2025}
}

🤝 Contributing

We welcome contributions! Please see our Contributing Guidelines for details.

Ways to Contribute

• Report bugs or dataset issues
• Suggest improvements to the generation pipeline
• Add new domain coverage
• Improve baseline models
• Submit evaluation results

📄 License

This dataset is released under the MIT License. Please ensure compliance with source data licensing when using.

🙏 Acknowledgments

• Data sources: Wikipedia, Britannica, Stanford Encyclopedia of Philosophy, Microsoft WikiQA, ArnetMiner, CNN, Daily Mail
• Built using: SpaCy, BERT, RoBERTa, Sentence-BERT, YAKE
• Human evaluation conducted via Amazon Mechanical Turk

📞 Contact

• Rohit Kumar: srz248004@iitd.ac.in
• Niladri Chatterjee:
• Institution: Indian Institute of Technology Delhi

🔗 Related Work

• TruthfulQA
• FEVER Dataset
• SciFact
• HaDeS

---

⭐ Star this repository if you find it useful for your research!