More Transparent, Safer, and Compact: Promoting Responsible AI with Gemma. Through the use of Gemma 2-powered content classifiers, ShieldGemma ensures safer AI interactions by removing potentially dangerous inputs and outputs.
The new best-in-class open models, Gemma 2, were made available in June in parameter sizes of 27 billion (27B) and 9 billion (9B). Since its release, the 27B model has risen to the top of the LMSYS Chatbot Arena leaderboard, surpassing well-known models by more than twice its size in real-world interactions.
However, Gemma is more than simply a show. It prioritizes accessibility and safety and is based on responsible AI. There are thrilled to present three new Gemma 2 family members in support of this commitment:
- Gemma 2 2B: A new iteration of this well-liked 2 billion (2B) parameter model, Gemma 2 2B offers a potent balance between performance and efficiency along with integrated safety improvements.
- ShieldGemma: a collection of safety content classifier models that filter AI model input and outputs to protect users, based on Gemma 2.
- Gemma Scope: Gemma Scope is a brand-new tool for model interpretability that provides unmatched visibility into the inner workings of these models.
With these enhancements, researchers and developers can now produce safer user experiences, acquire previously unheard-of insights into the models, and safely implement potent AI responsibly, directly on the device, opening up new avenues for creativity.
Gemma 2 2B: Experience Next-Gen Performance, Now On-Device
The Gemma 2 2B model, a much-anticipated addition to the Gemma 2 family, is now available. By using distillation to learn from larger models, this lightweight model generates disproportionately high results. Gemma 2 2B actually outperforms all GPT-3.5 models on the Chatbot Arena, showcasing its remarkable conversational AI capabilities.
Gemma 2 2B provides:
- Outstanding performance: Outperforms other open models in its category and offers the finest performance for its size.
- Cost-effective and adaptable deployment: Run Gemma 2 2B effectively on a variety of hardware, including laptops, edge devices, and reliable cloud installations using Google Kubernetes Engine (GKE) and Vertex AI. It is available as an NVIDIA NIM and is optimised with the NVIDIA TensorRT-LLM library to further increase its speed. Data centres, the cloud, local workstations, PCs, and edge devices with NVIDIA RTX, NVIDIA GeForce RTX GPUs, or NVIDIA Jetson modules for edge AI are among the deployments that this optimisation is intended for. For more efficient development, Gemma 2 2B also easily interacts with Keras, JAX, Hugging Face, NVIDIA NeMo, Ollama, Gemma.cpp, and soon MediaPipe.
- Accessible and open: accessible for research and commercial applications under the Gemma terms, which are advantageous for businesses. It’s even small enough to operate on Google Colab’s free T4 GPU tier, which makes development and experimentation simpler than ever.
The model weights for Gemma 2 are now available for download on Hugging Face, Vertex AI Model Garden, and Kaggle. Its skills can also be tested in Google AI Studio.
ShieldGemma: Protecting Users with State-of-the-Art Safety Classifiers
Developers and researchers must put in a lot of work to deploy open models ethically in order to guarantee inclusive, safe, and engaging AI outputs. Currently launching ShieldGemma, a line of cutting-edge safety classifiers made to identify and reduce hazardous content in AI model inputs and outputs, to assist developers in this process. Four major areas of harm are especially targeted by ShieldGemma:
- Hate speech
- The act of harassing
- Content that is sexually explicit
- Content that is dangerous
In addition to the Google Cloud off-the-shelf classifiers that are now available via API, these open classifiers enhance the current suite of safety classifiers in the Responsible AI Toolkit, which provides a methodology to develop classifiers tailored to a certain policy with a limited number of datapoints.
ShieldGemma can assist you in developing safer, more effective AI applications in the following ways:
- Performance of SOTA: ShieldGemma is the industry-leading safety classifier, built on top of Gemma 2.
- Variable sizes: ShieldGemma provides a range of model sizes to accommodate different requirements. While the 9B and 27B models offer better performance for offline applications where latency is less of an issue, the 2B model is best suited for online classification jobs. NVIDIA speed optimizations are used by all sizes to achieve effective hardware performance.
- Open and cooperative: ShieldGemma’s open architecture promotes openness and cooperation among AI experts, helping to shape the safety standards of the ML sector going forward.
Gemma Scope: Illuminating AI Decision-Making with Open Sparse Autoencoders
Researchers and developers now have unmatched insight into those Gemma 2 models’ decision-making processes to Gemma Scope. Gemma Scope functions as a powerful microscope, using sparse autoencoders (SAEs) to magnify particular areas of the model and make its internal mechanisms easier to understand.
The dense, complex data collected by Gemma 2 is broken down into a more easily analyzed and comprehendible format by these SAEs, which are specialised neural networks. Researchers can learn a great deal about Gemma 2’s pattern recognition, information processing, and prediction-making processes by examining these enlarged images. The goal with Gemma Scope is to assist the AI research community in learning how to create AI systems that are more comprehensible, accountable, and trustworthy.
What makes Gemma Scope revolutionary is this:
- Open SAEs: More than 400 openly accessible SAEs that cover all Gemma 2 2B and 9B layers.
- Interactive demos: Use Neuronpedia to examine SAE characteristics and examine model behaviour without writing code.
- User-friendly repository: Examples and code for interacting with Gemma 2 and SAEs.
A Future Based on Conscientious AI
These releases demonstrate this continued dedication to giving the AI community the instruments and materials required to create a future in which AI serves the interests of everybody. To think that creating safe and useful AI requires open access, transparency, and cooperation.
Start Now
- Download or test Gemma 2 2B using Google AI Studio or NVIDIA NIM to see its strength and effectiveness.
- Investigate ShieldGemma and create AI apps that are safer.
- To learn more about the inner workings of Gemma 2, try Gemma Scope on Neuronpedia.
