How did you make the chatbot stay in character so well?

#1
by VitalContribution - opened

Hey HappyDrone!

I tried out a few DialoGPT personalities but some of them changed there character after some messages.
Your Nexus Nova (sometimes repeats itself) but pretty much stays in character.
I would love get more information on how you created the this chatbots personality so well.

Best Regards
Vital (my discord: VitalContribution#2191)

Happy to hear that! I am not sure how others train their DialoGPT models, but I personally went with an approach of making my own dataset from scratch with the help of ChatGPT to speed up the process. I made about 1100 lines of dialogue back and forth between the potential "customer" and "salesman". I have 1000 lines of text only on the sale of ships and there is lots of repetition to try to get the AI to be consistent. I trained using the guide using collar https://github.com/RuolinZheng08/twewy-discord-chatbot. I think my settings were either 6 or 9 epochs, and self.per_gpu_train_batch_size = 1, self.per_gpu_eval_batch_size = 14. Don't quote me on that though, could have been different.

personally I was actually unhappy with the results so far but I am glad to hear you think otherwise. My main issue so far has been that the bot doesnt have the coherence I want it to. Moving around my prompts order in the training file might help with that, but I think it's a limitation of this model overall that might not be able to be overcome. One second its talking about one ship, but then the next second I ask it for more info, it will sometimes mention an entirely different ship. This doesnt always happen, but it's my biggest gripe with the model overall. I have been working on another version which broadens up the conversation topics away from ships. I might also try to train on the large model instead of medium, but havent had the time to compare pricing on cloud GPUs.

Me and my friends are making a collaborative sci fi D&D game where we all take turns DMing and I've created this bot as an experiment in creating NPCs that can talk on their own accord. My end goal is to place Nex and many others into my in universe codex system, which is basically our own in universe internet. This is still early days but for something I put together in under a day, im pretty happy with the progress

I've added you on discord if you want to discuss further and share my training data: Fourth Dimension#3938
Thanks and have a great day

Sign up or log in to comment