Commit 97b6bb2f by Jonathan Thomas

New experimental prompt strategy to only show valid and invalid combinations of…

New experimental prompt strategy to only show valid and invalid combinations of behavior emojis, instead of example phrases. Works well with gpt-4o-mini and llama3-70b. Does not work great with gpt-3.5-turbo. Also added negative LLM unit tests checks for behaviors (FOLLOW and not ATTACK).
parent 5d539e10
Pipeline #13247 passed with stages
in 2 minutes 34 seconds