Evaluating Language Models
The Data Exchange with Ben Lorica
00:00
The Seven Metrics of Detoxification Toxicity
Helm uses seven metrics to measure the performance of its models. The key idea is that there could be trade-offs between those. We're using the perspective API to detect whether there's something toxic. And then the last metric is efficiency, which actually does look at the some information about the internal model.
Transcript
Play full episode
