Multimodal and Safer: Conscientious AI with Gemma
ShieldGemma is a set of safety content classifier models that Google launched last year. It is based on Gemma 2 and is intended to identify harmful content in the text inputs and outputs of AI models.They are thrilled to announce ShieldGemma 2, building on the foundation of responsible AI as it launch Gemma 3 today.
Based on Gemma 3, ShieldGemma 2 is a 4 billion (4B) parameter model that helps you create reliable datasets and models by comparing the safety of your synthetic and natural photos to important categories. Researchers and developers can now effortlessly reduce the possibility of hazardous content in their models across important areas of harm with this addition to the Gemma family of models:
- Content that is sexually explicit
- Content that is dangerous
- Abuse
It suggest utilising ShieldGemma 2 as an output filter for image generating systems or as an input filter for vision language models. Both artificial and natural photos can be processed with ShieldGemma.
What makes ShieldGemma 2 different?
Beyond text, multimodal models present new problems for picture safety training and comprehension, which is why ShieldGemma 2 is designed to react to a broad variety of subtle and varied imagery types.
They carefully selected training datasets of both synthetic and natural photos in order to develop a reliable image safety model, and Gemma 3 was fine-tuned to show excellent performance.It will be publishing a technical report that includes third-party benchmarks in addition to the comparison of safety standards to the following benchmarks.
ShieldGemma can assist you in creating safer AI image apps in the following ways:
- Flexibility: Modify the prompt template to suit your needs and upload any artificial or natural photos. Use your own GPU or Google Colab to fine-tune.
- Versatility: It is compatible with all Gemma 3 tools, including well-known frameworks like Transformers, JAX, Keras, Ollama, and others.
- Collaborative: ShieldGemma is open by nature and invites community partners to continue constructing inclusively as all work to advance industry safety standards.
It takes the entire community to deploy open models responsibly, and it anticipate investigating in the near future how ShieldGemma 2 might be made available in smaller sizes, in more harm regions, and in accordance with the multimodal ML Commons taxonomy.
How might developers utilise ShieldGemma 2 effectively?
Google wants academics and developers to use ShieldGemma 2 in a number of important ways to reduce hazardous content in AI applications. ShieldGemma 2, a safety content classifier model based on Gemma 3, is made especially to identify dangerous material in both artificial and organic photos.
This is an explanation of how Google recommends using it:
As an input filter for vision language models: ShieldGemma 2 can be used by developers to verify that images are acceptable to use before feeding them into vision language models. This makes it easier to guarantee that the models are handling suitable and safe visual data.
ShieldGemma 2 can also be used as an output filter for picture generating systems, evaluating the security of AI-generated images prior to their display to consumers. This aids in stopping the spread of offensive or dangerous content produced by these systems.
To create robust datasets and models: Researchers and developers can create safer and more dependable training data for their models by identifying and minimising the presence of dangerous information in image datasets with ShieldGemma 2.
Across major harm areas: ShieldGemma 2 is made especially to scan for harmful content in a variety of categories, including violence, dangerous content, and sexually explicit content. This enables programmers to specifically target and reduce these dangerous picture kinds in their programs.
Customisation and flexibility: ShieldGemma 2 allows developers to upload any natural or synthetic image and modify the prompt template to tailor the model to their own requirements. Additionally, they can use their own GPU or Google Colab to refine the model.
Integration with current tools: ShieldGemma 2 is flexible and compatible with all Gemma 3 tools, including well-known frameworks like Ollama, JAX, Keras, and Transformers. This facilitates the integration of ShieldGemma 2 into the workflows of developers who are already utilising these technologies.
Communal cooperation: Google encourages partners to assist in developing and enhancing ShieldGemma 2 because it sees the responsible implementation of open models as a communal endeavour. The goal of this cooperative strategy is to consistently advance industry safety standards.
Google also predicts that ShieldGemma 2 will be delivered in smaller sizes, cover more harm regions, and be in line with the multimodal ML Commons taxonomy, all of which point to continued attempts to increase the tool’s usefulness in reducing harmful content. To get started, researchers and developers are invited to investigate ShieldGemma 2 on the developer site and on other platforms such as Hugging Face, Ollama, and Google AI Studio.
