Disclaimer: I only trained the LoRA adapter on the original qat model, rather than having the memory to include QAT in the training itself,
but they still seem to hold well against quantization.
Trained on private multiturn conversational dataset, continued creative pretraining on a private dataset,
and a dash of reasoning-included data from larger sibling Gemma 4 31B for the sake of not forgetting their skeleton.
~25 hours on a 3090, QLoRA SFT with unsloth.
Gemini 3 Flash seems to appreciate their creativity, at least; scored them just as well as the Gemma 4 31B models I can run locally.
