Changed the default sampling method for Layla to Dynamic Temperature Sampling. See the paper: https://arxiv.org/abs/2309.02772
This sampling method is reported to provide more organic responses from models by various users in the local AI community. It has already been implemented in other popular frontends such as Oobabooga, KoboldCpp, etc., and is scheduled to be added to SillyTavern as well, so it as seen a fair number of real-world use-cases.
You can turn this feature off by setting "Dynamic Temperature Range" setting in the Advanced Settings page to 0.
New features:
added "dynamic temperature sampling"
added "min P" config to be used with dynamic temperature sampling
Improvements:
added better estimation of character loading times
you are now able to adjust the frequency of proactive messaging
Bug fixes:
fixed a bug where character backgrounds were not shown correctly in the character selection screen on wider phones
fixed a bug where your astrologer in My Horoscopes app sometimes outputs the {{user}} template
fixed a bug where second time loading was not reading the cache in some cases so was loading slowly
Live on Google Play
Live on the App Store
Comments