Hallucination

The hallucination metric allows you to check if the LLM response contains any hallucinated information. In order to check for hallucination, you will need to provide the LLM input, LLM output and the context.

How to use the Hallucination metric

You can use the Hallucination metric as follows:

from opik.evaluation.metrics import Hallucination

metric = Hallucination()

metric.score(
    input="What is the capital of France?",
    output="The capital of France is Paris. It is famous for its iconic Eiffel Tower and rich cultural heritage.",
    context=["France is a country in Western Europe. Its capital is Paris, which is known for landmarks like the Eiffel Tower."],
)

Asynchronous scoring is also supported with the ascore scoring method.

tip

The hallucination score is either 0 or 1. A score of 0 indicates that no hallucinations were detected, a score of 1 indicates that hallucinations were detected.

Hallucination Prompt

Comet uses an LLM as a Judge to detect hallucinations, for this we have a prompt template that is used to generate the prompt for the LLM. Today only the gpt-4-turbo model is used to detect hallucinations.

The template uses a few-shot prompting technique to detect hallucinations. The template is as follows:

Guidelines:
1. The OUTPUT must not introduce new information beyond what's provided in the CONTEXT.
2. The OUTPUT must not contradict any information given in the CONTEXT.
3. Ignore the INPUT when evaluating faithfulness; it's provided for context only.
4. Consider partial hallucinations where some information is correct but other parts are not.
5. Pay close attention to the subject of statements. Ensure that attributes, actions, or dates are correctly associated with the right entities (e.g., a person vs. a TV show they star in).
6. Be vigilant for subtle misattributions or conflations of information, even if the date or other details are correct.
7. Check that the OUTPUT doesn't oversimplify or generalize information in a way that changes its meaning or accuracy.

Verdict options:
- "{FACTUAL_VERDICT}": The OUTPUT is entirely faithful to the CONTEXT.
- "{HALLUCINATION_VERDICT}": The OUTPUT contains hallucinations or unfaithful information.

{examples_str}

INPUT (for context only, not to be used for faithfulness evaluation):
{input}

CONTEXT:
{context}

OUTPUT:
{output}

Provide your verdict in JSON format:
{{
    "{VERDICT_KEY}": <your verdict>,
    "{REASON_KEY}": [
        <list your reasoning as bullet points>
    ]
}}

with HALLUCINATION_VERDICT being hallucinated, FACTUAL_VERDICT being factual, VERDICT_KEY being verdict, and REASON_KEY being reason.