replaces Gemma-4-31B-FT-it?
According to this message you posted https://huggingface.co/mradermacher/model_requests/discussions/2172#69d7204f89d95626b6ab9756
I'd take that aifeifei798/Gemma-4-31B-FT-it is now obsolete and this DarkIdol version is the fixed one? And this one also have extra fine tuning on top of that?
It's not exactly that the previous one is "obsolete," but rather a matter of model limitations.
In this release, Google integrated safety alignments across 39 layers. While my current fine-tuning method works exceptionally well for single-turn interactions, the "intelligence" and consistency of the model tend to degrade significantly during multi-turn conversations.
Since my primary focus is on Roleplay (RP) models—which demand high stability and logic over long-term dialogues—I’ve found that the Gemma 4 base isn't the ideal foundation for that specific goal. As a result, DarkIdol-Gemma-4-31B-it-FT will be my final optimization for this model series. It is the "fixed" and most polished version I can offer for this architecture.
As expected of Google, a pity, really. Thank you for the explanation, that was a really good answer.