Highlights for this update are improvements to the RMI app (Raw Model Instructions)!
You are now able to adjust prompt templates used for your custom models. ChatML and Vicuna are given as examples, but the configuration should really support any format.*
Also in the RMI app, you can manage storage and delete the custom models copied to your app's cache. This should help those of you who like to rapidly switch between lots of different models.
In answer to all your feedback regarding proactive messaging, "Proactive Chat" has been moved to its own app. This gives you fine-grained control over how you want it to work. You can select the frequency as usual, in addition, you can also choose which characters are allowed to proactively chat with you. Removing the "Proactive Chat" app stops all future notifications from being sent to you; adding it back re-enables the notifications. I hope this addresses the majority of use-cases!
Note: In this update proactive chat has been turned off for everyone. If you would like to re-enable it, please add the app to Layla in the App Screen.
Lastly, this version continues to lay the groundwork for long term memory, including a locally implemented vector database and embedding model with sub-1-second performance.
New features:
added "Proactive Chat" app
Improvements:
added "show performance counters" in advanced settings: this will show CPU, RAM usage and other system information during chatting
added ability to manage model storage space in Raw Model Instructions app
added details screen for each app, with a longer description on what the app does and instructions on how to use
added ability define custom prompt formats in RMI
Bug fixes:
fixed bug in advanced settings where you were unable to turn off dynamic temperature sampling
fixed bug where Precise settings were not turning off dynamic temperature
fixed bug where scheduled proactive messages were not cancelled even though proactive messaging is turned off
* Note: Layla expects a dynamic prompt template, and the characters are built with this in mind, so using a "fixed" prompt format such as Vicuna may give unexpected results. For example, Vicuna prompts have a hardcoded "ASSISTANT" prompt beginning, which may cause characters to refer to themselves an AI assistants.
From personal experience, I have found Layla's prompt format actually gives better results for chatting with characters and apps like Roleplay. For precise assistant features such as asking about facts and trivia, using your custom model's prompt format will give better results.
Comments