Tool Tutorial · Xiaohu Explains

Google ships the gemma-trainer skill: let AI fine-tune your own Gemma on your own machine

An open-source training playbook that locks down supervised fine-tuning, preference alignment, and reward scoring — plus LoRA parameters — paired with Unsloth, so a consumer GPU with 8GB of VRAM is enough to get started.
60-second briefing
  • Google added a skill called gemma-trainer to its open-source gemma-skills repo, built to teach AI assistants how to fine-tune Gemma models locally on your own machine.
  • It's essentially a playbook written for AI agents to read (SKILL.md), plus 5 ready-made Python scripts. You describe what you want in plain language, and the assistant picks parameters and runs the training.
  • It covers three training paths: SFT for teaching new knowledge, DPO for tuning style preferences, and RM for training a scoring model — each with its data format and key parameters locked in.
  • It defaults to Unsloth, which Unsloth claims cuts VRAM use by up to 70% and runs 2x faster — enough for a single consumer GPU to fine-tune 12B models or larger, including image and audio training.
  • Once trained, models can be converted to lightweight formats like GGUF and run on phones and IoT devices via LiteRT-LM.
⚑This is Google's own introduction of its tool, published on the DEV community. The performance figures (70% VRAM savings, 2x speed) are Unsloth's own claims. The mechanism details below come from the code in the open-source repo google-gemma/gemma-skills.
What you can do

Fine-tune a Gemma that only speaks your niche, on your own machine

You want a model of your own — one that knows the jargon of your obscure field, matches your tone, and only handles your specific job. Open Gemma was built for exactly this kind of customization, but actually fine-tuning it has always meant wading through a maze of environment setup and unreadable guides first.

Google's Gemma team recently smoothed out that first step: a new skill called gemma-trainer, added to the open-source gemma-skills repo.

With it, you can tell your AI assistant something like "fine-tune Gemma 4 E2B on this dataset," and it knows to validate your data first, pick the right LoRA parameters, kick off training with resource-friendly defaults, and evaluate the results. The whole local fine-tuning workflow becomes a playbook the assistant can read and execute directly.
💡Why it matters: it locks the whole decision tree for local Gemma fine-tuning — which training method to use, how to configure LoRA parameters, how to avoid blowing out your VRAM, how to convert the result for on-device deployment — into a single skill file. Paired with Unsloth, a consumer GPU with 8GB of VRAM is enough to get started, and it covers text, image, and audio training all at once.
📦 Open-source repo · Free
github.com/google-gemma/gemma-skills
What it actually is

It's essentially a playbook written for AI

gemma-trainer's core is a SKILL.md file plus 5 ready-made Python script templates. It's not an app, and it's not a training library you're meant to call directly yourself.

SKILL.md is a blueprint written specifically for AI agents to read. It locks down the best practices for local fine-tuning — what value each parameter should take, which pitfalls to avoid — all in one place. Once your assistant has read it, it goes from "roughly knowing how fine-tuning works" to "knowing exactly which script to adjust on this machine, and which knob to turn to what setting." The 5 scripts in the repo each handle one job: one runs supervised fine-tuning, one runs preference alignment, one trains a scoring model, one validates data formatting, and one handles data distillation.

You describe it"Fine-tune Gemma 4 E2B"
→
Agent reads
gemma-trainer
→
Picks the right script
sets parameters
→
Runs training
→
Evaluates
and iterates

The skill isn't tied to any particular AI tool — it works with anything from Gemini to Claude, and also works with Google's own Antigravity CLI tool (agy). Setup is a one-liner: copy the gemma-trainer folder into your AI assistant's skills directory, and it immediately knows how to guide you.

What's more interesting is that the assistant doesn't just blindly execute — it also catches your mistakes based on the capability constraints written into the skill. One real example: a training request accidentally specified Gemma 4 31B for an audio task, and the assistant flagged it on the spot, pointing out that 31B only handles text and images and has no audio capability, suggesting E2B or 12B instead.

Agent flags that Gemma 4 31B has no audio capability and suggests switching to E2B or 12B
If you ask for audio fine-tuning on a model with no audio capability, the assistant will catch it and offer viable alternatives. Source: DEV / Google AI
Three training paths