Matrixfag here. While we've been all over the place trying to fix bugs in our notebook (and there's still a lot of them), we're also working in the background continually training new 6B models in an effort to improve on our initial release. However, we don't want to suddenly replace our models with the newer ones without warning, nor do we want to be silent about what we're doing to make the bot better. To that end, we've introduced a "Pygmalion 6B Beta" option in our Colab notebook which allows you to test out our latest models to see if they're better or not. After all, the best judge for whether our models are good or not are not automated evaluations, nor even ourselves, but you guys. So, we invite you to come test our new model out for yourselves, and tell us what you think of it. To access it, simply go to the Colab notebook and select from the dropdown "Pygmalion-6B Beta". The notebook will do the rest.
Picrel is a log of what we've done on this run compared to our initial model. In only two days since our first release of 6B, we've received over 200 MB of CAI data and close to 5,000 downloads on HuggingFace - an absolutely tremendous amount of support! Of course, we're always accepting new CAI data, where a guide to doing that can be found at
https://rentry.org/chatlog-dumping (we changed the URL). If you were on the fence about submitting more data because you already submitted some and worry duplicated conversations would show up in the data, worry not. We've implemented a deduplicator on our end, meaning that only new conversations will be added to the training data, while your old conversations remain unchanged. In addition, as said in the log, we've fed a tiny bit of the SODA dataset into the model for the purposes of increasing the amount of SFW data we have in our set. To be clear: we are not trying to reduce the capabilities of our model to do NSFW, but rather to allow for more stable SFW conversations alongside the NSFW ones.
Your feedback is definitely being heard, and we're aware of all the problems our model still has. We're also very aware the notebook is very glitchy and barebones. Believe us, we're not fans of Gradio either. To that end, we're starting to work on a much fancier solution which will provide a much better chat experience, along with the opportunity to provide data (if you wish) for RLHF. It may take some time to do that, but it'll be worth the wait. Promise. As always, thank you all for the support you're showing. We've reached the #10 spot on the most downloaded (in the last month) conversational model on HuggingFace in only two days, only beaten out by models made by Facebook and Microsoft themselves! Absolutely insane. As usual, I'll be around to answer questions. Let us know what you think of the beta model!
(We've also heard your requests for saving chats. We're trying to get that out as soon as possible, but the way Gradio is structured, it's gonna take a little bit of time to get that in there. We're really sorry about that, but I promise, it is coming.)
