Measuring Hallucinations in RAG Systems

Source: https://vectara.com/measuring-hallucinations-in-rag-systems/ Publisher: Vectara · Published: 6 November 2023 · Author: Shane Connelly

Summary

Introduces Vectara’s Hallucination Evaluation Model (HEM) and the accompanying public Hallucination Leaderboard. The measured question is narrow and therefore tractable: given retrieved evidence, does the model “only use the data provided to it in generating its output”?

Leaderboard Results

ModelAnswer rateAccuracyHallucination rateAvg summary length
GPT-4100%97.0%3.0%81.1 words
GPT-3.599.6%96.5%3.5%84.1 words
Llama 2 70B99.9%94.9%5.1%84.9 words
Llama 2 13B99.8%94.1%5.9%82.1 words
Llama 2 7B99.6%94.4%5.6%119.9 words
Cohere-Chat98.0%92.5%7.5%74.4 words
Cohere99.8%91.5%8.5%59.8 words
Claude 299.3%91.5%8.5%87.5 words
Mistral 7B98.7%90.6%9.4%96.1 words
Google PaLM92.4%87.9%12.1%36.2 words
Google PaLM-Chat88.8%72.8%27.2%221.1 words

The Length Correlation

PaLM-Chat produced the longest summaries by a wide margin and hallucinated most — every extra sentence is another chance to exceed the evidence. Worth reading alongside the answer-rate column, which shows how often a model declined to answer at all.