Hi guys,
I’m still playing around with GPT-3 and trying to figure out some best practices. I’ve been creating a “clickworthy” blog title generator and I’ve been getting some good results but I absolutely want to improve where I can.
I’ve been providing DaVinci with about 400~ unique prompts and about 4000~ outputs - Is it a standard practice to provide duplicate prompts with unique outputs or should I be focusing on 1-in-1-out?
@hello3 if you’re using so many prompts, I highly suggest you finetune (a smaller model) instead of including the examples in the context. Prompted models don’t scale unless the prompt is small, and finetuned models perform much better and consistently.
I would also recommend against training on repititive inputs, as it makes the model brittle and repetitive (surprisingly). My observations might be use-case specific though, so you should always run an experiment and benchmark the two. You can also use the new edit model to mod your inputs a little so there’s a little more diversity.