[ggm]

Adversarial play might be interesting. I'm assuming the goal here is to use the feedback as training input, fitness to expectations and like.

So a random chance of a bad actor adversary? Or high scores for lying? "The traitor" style

Team models? I suppose I'm arguing for gamification with a leader board and hubristic bragging rights.

[flowingfocus]

Adversarial play sounds fun, I might try that out!

The LLM we use is mistral-large-latest, we didn't do any training on the data yet