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
| Model | Answer rate | Accuracy | Hallucination rate | Avg summary length |
|---|---|---|---|---|
| GPT-4 | 100% | 97.0% | 3.0% | 81.1 words |
| GPT-3.5 | 99.6% | 96.5% | 3.5% | 84.1 words |
| Llama 2 70B | 99.9% | 94.9% | 5.1% | 84.9 words |
| Llama 2 13B | 99.8% | 94.1% | 5.9% | 82.1 words |
| Llama 2 7B | 99.6% | 94.4% | 5.6% | 119.9 words |
| Cohere-Chat | 98.0% | 92.5% | 7.5% | 74.4 words |
| Cohere | 99.8% | 91.5% | 8.5% | 59.8 words |
| Claude 2 | 99.3% | 91.5% | 8.5% | 87.5 words |
| Mistral 7B | 98.7% | 90.6% | 9.4% | 96.1 words |
| Google PaLM | 92.4% | 87.9% | 12.1% | 36.2 words |
| Google PaLM-Chat | 88.8% | 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.
Related Concepts
- Hallucination Detection — primary topic
- RAG · LLM as Judge · Search Evaluation