Grokking LLM Evals
Vote

0% completed

​

Help Center Release Thresholds

Suppose a new search method finds the right article in 267 of the 300 answerable test cases, while the current one finds it in 246. Would you ship the new version?

Most teams would, and here they would ship a bot that answers more of the questions its articles do not cover. The better search finds loosely related chunks for those questions too, and the model then writes an answer from them. Only a gate on unanswerable questions catches this, and only if it was written before the run.

This lesson sets the release gates for the help center bot and checks how much each gate can detect

.....

.....

.....

Like the course? Get enrolled and start learning!

Reading Progress

0%


Vote for new content