Hallucination detection

[AndreaPi]

Detect hallucinations in LLM outputs

Details

• What is the problem you're trying to solve?
  • detect hallucinations in LLM outputs
• What tasks or workflows would be enabled by having support for your proposed feature in cleanlab?
  • selective classification, classification with abstention, and more in general a way to "trust" your LLM output
• Can you share code snippets or pseudocode describing uses of your feature?
  • I don't have code, but there are a few GitHub repositories implementing hallucination detection models, such as https://github.com/oneal2000/MIND
• Can you share any datasets that can help us assess the usefulness of the proposed feature?
  • https://github.com/oneal2000/MIND/tree/main/helm

[jwmueller]

Thank you for the suggestion. Note we do offer the Trustworthy Language Model, which is exactly designed for hallucination detection.

Relevant tutorials:
https://help.cleanlab.ai/tutorials/tlm/ https://help.cleanlab.ai/tutorials/tlm_custom_model/ https://help.cleanlab.ai/tutorials/tlm_rag/#alternate-low-latencystreaming-approach-use-tlm-to-assess-responses-from-an-existing-rag-system

Hallucination-detection benchmarks in RAG: https://towardsdatascience.com/benchmarking-hallucination-detection-methods-in-rag-6a03c555f063