Making LLMs hallucinate less, how to diagnose ML models, from PM in Google AI to CEO of Galileo - Vikram Chatterji - The Data Scientist Show #066
The Data Scientist Show - Daliana Liu
00:00
How to Evaluate LLMs That Aren't Working Very Well
There are quite a few things that you could do. One metric which we've worked on for a bit is called the Data Eddert Potential Score. This is especially useful for folks if you're familiar with NLP, people who are building our natural language processing models. That DEP score as we call it has been very effective in terms of finding the long tail of data that was pulling the model performance down. It can essentially look at how confused was the model as it was trying to converge towards a particular label.
Transcript
Play full episode
