Closed Angel-Sb closed 1 year ago
the language model has no perception of time.
Thanks for commenting. I have noticed that the code for time configuration is on image generation prompt file (function.cs) and that the language model doesn't have perception of time. My problem is that even tho everything seems normal code wise the actual result is a bit weird.
The time actually is written into the character prompt by default if you are using the default github code as it is now, but I'll say this: you will need a way better model than pygmalion for it to work.
I was able to get a fairly stable result of the LLM working out what time it was only by using a better model. The current time in japan is written into the character prompt but a lot of the LLMs out there just simply do not listen to it and speak a random time. If you can get your hands on a good 13B model it seems to solve the issue, but it may also be possible with a high quality 7B model.
I would recommend checking out a few models and seeing what results you get. Check online what the best model is nowadays,
Thank you a lot! I will download and try different models to see what works best as you said.
I have tried to add a new time zone to the code.
Every time i ask sally for the time she couldn't be any more random.
Everything else seems to be working perfectly fine both oobabooga and stable diffusion. (The model that i use for oobabooga is mayaeary_pygmalion-6b_dev-4bit-128g)
Any help really appreciated!