gemma-4-12b-auto-v4

Gemma 12B Auto v4

This is the first proper release in an experimental fine-tuning series based on unsloth/gemma-4-12b-it.

Training started from the General Inquiry Thinking / Chain-of-Thought dataset and was supplemented with custom repair and broad-coverage examples created for this project. The goal is to help the model learn when reasoning is useful and when a short, direct answer is the better choice.

Small behavioral evaluation

We compared the model with the original Gemma 4 12B Instruct base on an internal suite of 40 fixed prompt/seed combinations. Both models were evaluated as Q6_K GGUFs with temperature 0.7, a maximum of 3,000 new tokens, and identical prompts and seeds.

Compared with the base model, Auto v4 produced final answers that were 68.1% shorter on average with Thinking enabled and 27.3% shorter in Instruct mode. With Thinking enabled, it answered directly in 35 of 40 cases while keeping a separate reasoning path available for selected tasks.

This is a small internal behavioral test focused on reasoning routing and verbosity, not a standardized benchmark or a comprehensive accuracy evaluation. The single empty, length-limited Thinking case came from a known ambiguous scheduling prompt shared by both runs.

The model was trained with Unsloth and Hugging Face TRL.

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>