I wanted a tiny decision model, so I built one with Gemma 270M

[rjn32s]

I wanted a tiny decision model.

Could Gemma 270M handle structured decisions locally instead of using a much larger LLM?

That became GemmaDecision.

It takes a situation + a set of choices and ranks them locally.

• Fine-tuned Gemma 270M
• CPU/GPU inference
• PydanticAI support
• Rust-powered HTTP serving
• Python package: gemmadecision

I also published the evaluation results and limitations:

https://github.com/rjn32s/GemmaDecision-270M

Would love to hear what you think or what you'd try with a 270M decision model.