My thoughts
Gemma-3 is very good, but it's very much ruined by censorship. Hell, I'm sure if they released Gemma-3 4B F16 and Gemma-3 12B, while removing all the censorship, maximizing its quality like you can with Imatrix, with output attachment tensors on F16 and other improvements, If you had further refined Gemma-3 on various better RPG datasets, it would make an incredible masterpiece. What do you say, are you planning on creating something like this? I'd really like to see something like that from you.
Excellent - I agree 100%.
Censorship hurts the model - both in terms of "pg rated" generation as well as "adult".
Qwen 32B was mostly uncensored at time of release.
Hopefully the next Llamas will follow their lead.
I am awaiting the decensor/alb versions of Gemma - problem is this damages the model.
There are other ways around it, but it is touch and go.
Frankly, I am hoping that Gemma 3 is much more merge friendly than Gemma2 ; which will help bring this model up to its true potential.
Dataset(s) + properly uncensored -> It is coming.
Excellent - I agree 100%.
Censorship hurts the model - both in terms of "pg rated" generation as well as "adult".Qwen 32B was mostly uncensored at time of release.
Hopefully the next Llamas will follow their lead.I am awaiting the decensor/alb versions of Gemma - problem is this damages the model.
There are other ways around it, but it is touch and go.Frankly, I am hoping that Gemma 3 is much more merge friendly than Gemma2 ; which will help bring this model up to its true potential.
Dataset(s) + properly uncensored -> It is coming.
About the data: this model PocketDoc/Dans-SakuraKaze-V1.0.0-12b is very, I would say very good. You can look at the dataset this author uses and if there is one, please consider it, because this model answers the way it should in terms of roleplaying, although it is not perfect.