About prompting for Mahjong

Hi there, i am doing a little project about using gpt3.5 turbo api to do text generation on Mahjong Games and i already finetune a personal gpt-3.5 turbo model for my work.

I will input the cards received by the player, along with their current hand and actions taken, as a prompt, and I hope to receive a conversation that fits the context. This conversation should not reveal information about my own hand or the cards received to other players.

The generated responses, while contextually appropriate, tend to inadvertently disclose my information to other players. Additionally, the tone lacks some resemblance to human conversation. I’m looking for any suggestions to address these issues. Thanks a lot!

Don’t always try to get everything to work flawlessly in one take like an elaborate OK Go music video.

You can edit.

Run the response back through the model and ask it if it reveals information about your cards and, if so, revise the text so it doesn’t.

You can also send the text (without hand revealing information) back through a model and ask it to be humanized, bantier, more humorous, whatever you want.

You can sometimes spend so much more time, effort, and money trying to get the response to be a perfect one-shot when it’s much quicker and easier to just revise.

1 Like