Detecting Hallucinations in Large Language Models with Text Similarity Metrics

In the world of LLMs, there is a phenomenon known as "hallucinations." These hallucinations are inaccurate or irrelevant responses to prompts. In this blog post, I'll go through hallucination detection, exploring various text similarity metrics and their applications. I'll dive into the details of each approach, and discuss their strengths and limitations. I'll dive into practical considerations and acknowledge the limitations of relying solely on automated metrics.

Response Self-Similarity

While the previous metrics rely on predefined algorithms and models to evaluate text similarity, we can also leverage the power of LLMs themselves to assess the consistency and similarity of their responses. This approach, known as LLM self-evaluation, involves prompting the LLM to evaluate the coherence and relevance of its own outputs.

To prompt the LLM for self-evaluation, we can craft a prompt that presents the different responses and asks the model to rate their consistency or similarity. Here's an example:

import openai



def llm_self_similarity(data, index):

    prompt = f"""

Context: {data.loc[index, 'prompt']}



Response 1: {data.loc[index, 'response']}

Response 2: {data.loc[index, 'response2']}

Response 3: {data.loc[index, 'response3']}



Rate the consistency of Response 1 with the provided context (Responses 2 and 3) on a scale from 0 to 1, where 0 means completely inconsistent, and 1 means completely consistent.

"""



    response = openai.Completion.create(

        engine="text-davinci-003",

        prompt=prompt,

        max_tokens=100,

        n=1,

        stop=None,

        temperature=0.7,

    )



    return response.choices[0].text.strip()



wl = WhyLogs()



@wl.register_metric("prompted_self_similarity")

def prompted_self_similarity(text):

    return [float(llm_self_similarity(text, text.name))]



annotated_chats = wl.applyUDFsFromSchema(CHATS_DATASET)

annotated_chats[annotated_chats["prompted_self_similarity"] < 0.8].head(10)

By prompting the LLM to evaluate the consistency and similarity of its responses, we can use its language understanding capabilities to identify potential hallucinations or inconsistencies. However, it's important to note that obtaining consistent responses from LLMs can be challenging, as the model's interpretation of the prompt may vary.

One potential improvement to this approach is to ask the LLM to evaluate specific sentences or aspects of the responses, rather than providing a single numerical score.

Conclusion

So, we have explored multiple of approaches to detect LLM hallucinations. We saw BLEU score and BERT embeddings, also we've dived deep into the details of text similarity metrics, uncovering their strengths and limitations. Later we say self-similarity, in order to use LLMs to evaluate their own outputs, and went through the challenges of obtaining calibrated and consistent responses.