About ShieldGemma
ShieldGemma is a set of instruction-tuned models for evaluating the safety of text and images against a defined set of safety policies. The ShieldGemma family includes open weights that allow users to tune the model for their specific use case.
The models can be used as part of a broader generative AI application deployment to help evaluate and prevent the likelihood of generative AI applications violating safety policies. ShieldGemma 2 is a 4 billion parameter model designed to label images for safety purposes. ShieldGemma 1 is based on Gemma 2 at parameter sizes of 2B, 9B and 27B.
Drafted from ai.google.dev · not yet confirmed by Google
How it works
ShieldGemma is a set of small classifiers from Google built on Gemma. You give the model a policy description and a piece of text (or, in the second version, an image) and it returns a yes or no on whether the content violates that policy, with a probability. The policies it was trained on cover sexually explicit content, dangerous content, harassment and hate speech, and it works reasonably well when you write your own.
It runs on a single GPU at the small sizes and is meant as an input and output filter around your own model. It's released under Gemma's terms, which are permissive but not a standard open source license.
From the project
Reviews
Be the first to review it
AnonymousKnow a team weighing ShieldGemma? Send them this page.