The Med-HALT benchmark calculates cosine similarity between model output embeddings and two references: Answer Similarity (between the correct option and model output) and Question Similarity (between the original question and model output).

Authors

• Person: Not available Organization: medRxiv
  Medical Hallucination in Foundation Models and Their ...

Sources

• Medical Hallucination in Foundation Models and Their ... www.medrxiv.org medRxiv via serper

Referenced by nodes (1)

• cosine similarity concept