2 ms·
It’d be nice if this post included a high-level cookbook for training the 3rd approach. The hand-waving around RL sounds great, but how do you accurately simula
by ed 2y ago
It’d be nice if this post included a high-level cookbook for training the 3rd approach. The hand-waving around RL sounds great, but how do you accurately simulate a customer for learning at scale?
- coredog64 2y agoAmazon Connect (as just one example) allows for post-call transcription (“calls may be recorded for quality assurance purposes”). Feed several thousand of these into an LLM alongside the ticket summary of initial problem and final outcome. With a month’s worth of calls you probably have a reasonable dataset to distill out a cheap 1B or 3B Llama model that is practically free compared to the real model used in your support agent workflow.