/vt/ - Virtual Youtubers

Matrixfag here. We've released a new version of our model on the dev branch of the repo, called V5. With this version, we've changed the way the model is trained. Instead of using "unsupervised fine tuning", where the bot learned how much it sucked based on the entire input prompt and response, we use "supervised fine tuning", which only tells the bot how much it sucked based on only the bot's response. It took us a fuckton of time to get it even kinda working properly, but we hope it works out. Picrel is a changelog which goes into further detail of everything changed from V4 to V5.

Worth noting that because of the different training method used, the same techniques and sampling settings used to boost the quality of output before may not work as well now. We highly encourage playing around with settings to see what works for you. The changelog in the picture shows a potential starting point for settings, but we've left the settings in the notebook as they are. Don't be afraid to report back any findings and feedback. We're definitely listening, especially since this is our first time using supervised fine tuning for the model. I'll be answering any questions, as per usual.

The changelog attached can also be found at https://github.com/PygmalionAI/logbooks/blob/master/2023-01-14.md.. We hope you enjoy the new dev model!