head of design
Gemma is designed with AI principles at the forefront. As part of making Gemma’s pre-trained models secure and reliable, we used automated techniques to exclude certain personal information and other sensitive data from the training set. Additionally, we used extensive fine-tuning from human feedback (RLHF) and reinforcement learning to align the instruction-tuned model to responsible behavior. To understand and mitigate the risk profile of the Gemma model, we conducted a robust assessment that included manual red teaming, automated adversarial testing, and evaluating the model’s capabilities against risky activities. A summary of these ratings is provided on the model card.
We are also releasing a new Responsible Generative AI Toolkit in collaboration with Gemma to help developers and researchers prioritize building safe and responsible AI applications. The toolkit includes:
Safety Classification: We provide a new methodology for building robust safety classifiers with minimal examples. Debugging: Model debugging tools help you examine Gemma’s behavior and address potential issues. Guidance: Access model builder best practices based on Google’s development experience. Deploy language models at scale.
Optimized across frameworks, tools, and hardware
You can fine-tune Gemma models based on your own data to adapt them to the needs of specific applications, such as summarization and search augmentation generation (RAG). Gemma supports a variety of tools and systems.
Multi-framework tools: Deploy your favorite frameworks with reference implementations for inference and fine-tuning across multi-framework Keras 3.0, native PyTorch, JAX, and Hugging Face Transformers. Cross-device compatibility: Gemma models run on common device types, including: Deliver broadly accessible AI capabilities on laptops, desktops, IoT, mobile, and the cloud. Cutting-edge hardware platform: We partnered with NVIDIA to optimize Gemma. NVIDIA GPUs ensure industry-leading performance and integration with cutting-edge technology, from the data center to the cloud to your local RTX AI PC. Optimized for Google Cloud: Vertex AI offers extensive tuning options and a comprehensive MLOps toolset with one click. Deployment using built-in inference optimization. Fully managed Vertex AI tools or self-managed GKE allow for advanced customization, including deployment from either platform to cost-effective infrastructure spanning GPUs, TPUs, and CPUs.
Free credits for research and development
Gemma is built for an open community of developers and researchers driving AI innovation. Get started with Gemma today with free access to Kaggle, a free tier of Colab notebooks, and a $300 credit for first-time Google Cloud users. Researchers can also apply for up to $500,000 in total Google Cloud credits to accelerate their projects.
Start
Find out more about Gemma and access the quickstart guide at ai.google.dev/gemma.
As we continue to expand our Gemma model family, we look forward to introducing new variants for a variety of applications. Stay tuned for events and opportunities in the coming weeks to connect, learn and build with Gemma.
I look forward to seeing what you create.
