In Loving memory of Simon Mark Hughes...

Quickstart: try HHEM-2.1-Open Live Demo

👉 Launch Interactive Demo - No setup required, runs in your browser

💡 Quick test: Try inputting "The capital of France is Berlin" as premise and "The capital of France is Paris" as hypothesis to see HHEM detect this factual but hallucinated case.

HHEM-2.1-Open is the latest open source version of Vectara's HHEM series models for detecting hallucinations in LLMs. These are particularly useful in the context of building retrieval-augmented-generation (RAG) applications or Agentic workflows, where a set of facts is summarized by an LLM, and HHEM can be used to measure the extent to which this summary is factually consistent with the facts.

Hallucination Detection 101

By "hallucinated" or "factually inconsistent", we mean that a text (hypothesis, to be judged) is not supported by another text (evidence/premise, given). You always need two pieces of text to determine whether a text is hallucinated or not. When applied to RAG or AI Agents, the LLM is provided with several pieces of text (often called facts or context) retrieved from some dataset, and a hallucination would indicate that the summary (hypothesis) is not supported by those facts (evidence).

A common type of RAG hallucination is factual but hallucinated. For example, given the premise "The capital of France is Berlin", the hypothesis "The capital of France is Paris" is hallucinated -- although it is true in the world knowledge. This happens when LLMs do not generate content based on the textual data provided to them as part of the RAG retrieval process, but rather generate content based on their pre-trained knowledge.

Additionally, hallucination detection is "asymmetric" or is not commutative. For example, the hypothesis "I visited Iowa" is considered hallucinated given the premise "I visited the United States", but the reverse is consistent.

💡 Using HHEM in production? We'd love to hear about your use case! Connect with us on LinkedIn or Twitter.

AI Engineers: Stay Updated on Hallucination Mitigation

Hallucinations mitigation is a growing field, that includes both hallucination detection and correction. Specifically:

• HHEM is a specialized model for detecting hallucinations in LLM outputs - it identifies when generated text is not supported by the provided context in RAG.
• VHC (Vectara Hallucination Corrector) can correct hallucinations - fix inaccurate generated content (for RAG or Agents), ensuring its consistent with the context. Learn more about VHC.

Together, these tools provide a comprehensive approach to hallucination mitigation, helping you to create better and more reliable AI applications.

Want to learn more? Here are some resources to learn more about hallucination mitigation and AI safety:

1. 📧 Join our newsletter for up to date updates on hallucination mitigation.

2. 🚀 Sign up for Vectara - Access Vectara Hallucination Corrector (VHC) and gain hands-on experience using RAG and Agent workflows with VHC.

3. Join our community:

• 🔗 LinkedIn: Follow @Vectara for AI safety insights and industry updates
• 🐦 X/Twitter: @vectara for real-time updates and research discussions
• ⭐ GitHub: Star our hallucination leaderboard which tracks LLM hallucination rates across leading LLMs.

Using HHEM-2.1-Open

Here we provide several ways to use HHEM-2.1-Open in the transformers library.

You may run into a warning message that "Token indices sequence length is longer than the specified maximum sequence length". Please ignore it which is inherited from the foundation, T5-base.

Using with AutoModel

This is the most end-to-end and out-of-the-box way to use HHEM-2.1-Open. It takes a list of pairs of (premise, hypothesis) as the input and returns a score between 0 and 1 for each pair where 0 means that the hypothesis is not evidenced at all by the premise and 1 means the hypothesis is fully supported by the premise.

from transformers import AutoModelForSequenceClassification

pairs = [ # Test data, List[Tuple[str, str]]
    ("The capital of France is Berlin.", "The capital of France is Paris."), # factual but hallucinated
    ('I am in California', 'I am in United States.'), # Consistent
    ('I am in United States', 'I am in California.'), # Hallucinated
    ("A person on a horse jumps over a broken down airplane.", "A person is outdoors, on a horse."),
    ("A boy is jumping on skateboard in the middle of a red bridge.", "The boy skates down the sidewalk on a red bridge"),
    ("A man with blond-hair, and a brown shirt drinking out of a public water fountain.", "A blond man wearing a brown shirt is reading a book."),
    ("Mark Wahlberg was a fan of Manny.", "Manny was a fan of Mark Wahlberg.")
]

# Step 1: Load the model
model = AutoModelForSequenceClassification.from_pretrained(
    'vectara/hallucination_evaluation_model', trust_remote_code=True)

# Step 2: Use the model to predict
model.predict(pairs) # note the predict() method. Do not do model(pairs). 
# tensor([0.0111, 0.6474, 0.1290, 0.8969, 0.1846, 0.0050, 0.0543])

Using with pipeline

In the popular pipeline class of the transformers library, you have to manually prepare the data using the prompt template in which we trained the model. HHEM-2.1-Open has two output neurons, corresponding to the labels hallucinated and consistent respectively. In the example below, we will ask pipeline to return the scores for both labels (by setting top_k=None, formerly return_all_scores=True) and then extract the score for the consistent label.

from transformers import pipeline, AutoTokenizer

pairs = [ # Test data, List[Tuple[str, str]]
    ("The capital of France is Berlin.", "The capital of France is Paris."),
    ('I am in California', 'I am in United States.'),
    ('I am in United States', 'I am in California.'),
    ("A person on a horse jumps over a broken down airplane.", "A person is outdoors, on a horse."),
    ("A boy is jumping on skateboard in the middle of a red bridge.", "The boy skates down the sidewalk on a red bridge"),
    ("A man with blond-hair, and a brown shirt drinking out of a public water fountain.", "A blond man wearing a brown shirt is reading a book."),
    ("Mark Wahlberg was a fan of Manny.", "Manny was a fan of Mark Wahlberg.")
]

# Prompt the pairs
prompt = "<pad> Determine if the hypothesis is true given the premise?\n\nPremise: {text1}\n\nHypothesis: {text2}"
input_pairs = [prompt.format(text1=pair[0], text2=pair[1]) for pair in pairs]

# Use text-classification pipeline to predict
classifier = pipeline(
            "text-classification",
            model='vectara/hallucination_evaluation_model',
            tokenizer=AutoTokenizer.from_pretrained('google/flan-t5-base'),
            trust_remote_code=True
        )
full_scores = classifier(input_pairs, top_k=None) # List[List[Dict[str, float]]]

# Optional: Extract the scores for the 'consistent' label
simple_scores = [score_dict['score'] for score_for_both_labels in full_scores for score_dict in score_for_both_labels if score_dict['label'] == 'consistent']

print(simple_scores)
# Expected output: [0.011061512865126133, 0.6473632454872131, 0.1290171593427658, 0.8969419002532959, 0.18462494015693665, 0.005031010136008263, 0.05432349815964699]

Of course, with pipeline, you can also get the most likely label, or the label with the highest score, by setting top_k=1.

HHEM-2.3 and the LLM Hallucination Leaderboard

See how LLMs compare: HHEM-2.3, our latest commercial hallucination detection model, powers our live leaderboard that continuously benchmarks leading LLMs for hallucination rates. Watch comparisons of the latest models including GPT-5, Claude 4, Grok-4, Gemini 2.5, Llama4, Mistrak, DeepSeek, and many others.

HHEM-2.3 advantages over the open source version:

• Enhanced accuracy and performance
• Cross-lingual support (English, German, French, Portuguese, Spanish, Arabic, Chinese-Simplified, Korean, Russian, Japanese, and Hindi)
• Available exclusively via Vectara's platform

🔍 Track AI safety across the industry: The leaderboard updates regularly as new models are released, giving the community insights into which LLMs are most reliable for factual tasks. All data and methodology are open source.

HHEM-2.1-Open vs. HHEM-1.0

HHEM-1.0 is the first open version of HHEM. The major difference between HHEM-2.1-Open and the original HHEM-1.0 is that HHEM-2.1-Open has an unlimited context length, while HHEM-1.0 is capped at 512 tokens. The longer context length allows HHEM-2.1-Open to provide more accurate hallucination detection for RAG which often needs more than 512 tokens.

The tables below compare the two models on the AggreFact and RAGTruth benchmarks, as well as GPT-3.5-Turbo and GPT-4. In particular, on AggreFact, we focus on its SOTA subset (denoted as AggreFact-SOTA) which contains summaries generated by Google's T5, Meta's BART, and Google's Pegasus, which are the three latest models in the AggreFact benchmark. The results on RAGTruth's summarization (denoted as RAGTruth-Summ) and QA (denoted as RAGTruth-QA) subsets are reported separately. The GPT-3.5-Turbo and GPT-4 versions are 01-25 and 06-13 respectively. The zero-shot results of the two GPT models were obtained using the prompt template in this paper.

Table 1: Performance on AggreFact-SOTA

Table 2: Performance on RAGTruth-Summ

Table 3: Performance on RAGTruth-QA

The tables above show that HHEM-2.1-Open has a significant improvement over HHEM-1.0 in the RAGTruth-Summ and RAGTruth-QA benchmarks, while it has a slight decrease in the AggreFact-SOTA benchmark. However, when interpreting these results, please note that AggreFact-SOTA is evaluated on relatively older types of LLMs:

• LLMs in AggreFact-SOTA: T5, BART, and Pegasus;
• LLMs in RAGTruth: GPT-4-0613, GPT-3.5-turbo-0613, Llama-2-7B/13B/70B-chat, and Mistral-7B-instruct.

HHEM-2.1-Open vs. GPT-3.5-Turbo and GPT-4

From the tables above we can also conclude that HHEM-2.1-Open outperforms both GPT-3.5-Turbo and GPT-4 in all three benchmarks. The quantitative advantage of HHEM-2.1-Open over GPT-3.5-Turbo and GPT-4 is summarized in Table 4 below.

Table 4: Percentage points of HHEM-2.1-Open's balanced accuracies over GPT-3.5-Turbo and GPT-4

Another advantage of HHEM-2.1-Open is its efficiency. HHEM-2.1-Open can be run on consumer-grade hardware, occupying less than 600MB RAM space at 32-bit precision and elapsing around 1.5 second for a 2k-token input on a modern x86 CPU.

Hallucination detection with Vectara

Vectara provides a Trusted Generative AI platform. The platform allows organizations to rapidly create an AI agent experience which is grounded in the data, documents, and knowledge that they have. Vectara solves critical problems required for enterprise adoption of RAG and Agentic AI applications, namely: reduces hallucination, provides explainability / provenance, enforces access control, allows for real-time updatability of the knowledge, and mitigates intellectual property / bias concerns from large language models.

HHEM-2.3 is fully integrated into Vectara and is automtically returned with every query API call.

To start benefiting from HHEM-2.3, you can sign up for a Vectara account, and you will get the HHEM-2.3 score returned with every query automatically.

Here are some additional resources:

1. Vectara API documentation.
2. Vectara SDK.
3. Learn more about Vectara's Boomerang embedding model, Slingshot reranker, and Mockingbird LLM

Cite this model

@misc {hhem-2.1-open,
	author       = {Miaoran Li and Rogger Luo and Ofer Mendelevitch},
	title        = {{HHEM-2.1-Open}},
	year         = 2024,
	url          = { https://huggingface.co/vectara/hallucination_evaluation_model },
	doi          = { 10.57967/hf/3240 },
	publisher    = { Hugging Face }
}