Gemma-3-Fornax-V4-27B-QAT

<p align="left"> <img width="65%" src="Fornax.jpg"> </p>

Gemma 3 27B V4 Fornax

Gemma Fornax is a distillation of the updated R1 05/28 onto Gemma 3 27B, with a particualar focus on timely and generalizable reasoning beyond coding and math. Most other open source thinking models, especially on the smaller side, fail to generalize their reasoning to tasks other than coding or math due to an overly large focus on GRPO zero for CoT which only generalizes for coding and math.

Instead of using GRPO, this model aims to SFT a wide variety of high quality, diverse reasoning traces from Deepseek R1 05/28 onto Gemma 3 to force the model to learn to effectively generalize its reasoning capabilites to a large number of tasks as an extension of the LiMO paper's approach to Math/Coding CoT.

Varying CoT length in conjuction with explicit noise regularization during training also prevents the characteristic length overfitting of GRPO, which tends to manifest as waffling, where the model reasons to a set length even when it has already reached an answer.

Recommended Settings

Temp .7 + Nsigma 1

Special Thanks:

Google for open sourcing the excellent Gemma 3 model line.