Everything K-culture — comebacks to K-beauty, straight to your inboxGet it in your inbox

METAL MEDIA

google-gemma/gemma-skills

918PythonApache-2.0

Google's Gemma team packages ready-made instructions so AI coding assistants can build and train Gemma models correctly

gemma-skills is a small repository of 'skill' files that teach AI agents how to work with Google's Gemma models. It currently offers two skills: one for building applications with Gemma, and one for fine-tuning or training Gemma on local hardware. The skills can be installed through third-party CLI tools like Vercel's skills.sh or Context7.

What it does

  1. Provides two skills so far: gemma-dev for building apps with Gemma or answering general Gemma questions, and gemma-trainer for fine-tuning Gemma using methods like SFT, DPO, RLHF, or reward modeling on local hardware.
  2. Skills are installed via external CLI tools (npx skills add or npx ctx7 skills install), not by manually copying code, letting an AI coding agent pull in the right guidance automatically.
  3. The project explicitly states it is not an officially supported Google product and is not covered by Google's Open Source Software Vulnerability Rewards Program.
  4. The repository itself is mostly documentation and configuration for these skills rather than a full application, aimed at making AI agents more reliable when handling Gemma-related tasks.

Why it matters

As AI coding agents increasingly write and fine-tune models on their own, having a standardized 'skill' package reduces the chance of agents giving wrong or outdated guidance about a specific model family like Gemma. For developers building or training Gemma models, this offers a shortcut to correct, maintained instructions instead of piecing together scattered tutorials.

Terms in this repo

  • SFT · Supervised Fine-Tuning, adapting a pretrained model using labeled example data
  • DPO · Direct Preference Optimization, a method to tune models based on preferred vs. non-preferred responses
  • RLHF · Reinforcement Learning from Human Feedback, training a model using human ratings as reward signals
  • Reward Modeling · Building a separate model that scores outputs, used to guide further training
  • skills CLI · A command-line tool that lets users browse and install predefined 'skill' packages for AI agents

Repository description (English)

Skills for the Gemma and model/agent interactions

Owner · google-deepmind

Open on GitHub

Trending repos

All repos →

Latest from METAL MEDIA