Google Developers has reportedly released Gemma 2 in June, having 27B and 9B parameter sizes. Since its existence, this model has become one of the highest-ranking open models on the LMSYS Chatbot Arena leaderboard. Following its advanced performance ‘Gemma’ is built on a foundation of responsible AI, which prioritizes safety and accessibility. This model also comes with three latest editions to the Gemma 2 family:
Google Offers Three New Models Of Gemma 2 Family
- Gemma 2 2B: a brand new version of the 2 billion (2B) parameter model, featuring built-in safety advancements and a powerful balance of performance and efficiency.
- ShieldGemma: to keep the user safe and to filter the input and outputs of AI models, focuses on a classifier model of safety contents built upon Gemma 2.
- Gemma Scope: a tool that interprets the offers of unparalleled insight into their model’s inner workings.
These new features enable researchers and developers to create a safer customer experience by utilizing a powerful AI responsibility.
Gemma 2 2B: Experience next-gen performance, now on-device
This lightweight model produces outsized results by learning from large models through distillation. Also, the Gemma 2 2B conversational AI ability suppresses the GPT 3.5 models on Chatbot Arena.
- Exceptional performance: delivers best-in-class performance for its size as compared to other models in the category.
- Flexible and cost-effective deployment: Gemma 2 2B can be run on hardware devices and boost cloud deployments with vertex AI and Google KubernetesEngine (GKE). its speed is optimized as ‘NVIDIA TensorRT-LLM library and available as ‘NVIDIA NIM. This optimization targets various deployments and seamlessly integrates with Keras, JAX, Hugging Face, and others.
- Open and accessible: available under the commercially-friendly Gemma terms for research and commercial application, also it’s small enough to run on the free tier of T4 GPUs in Google Colab. Its capabilities are also be tried in Google AI Studio.
ShieldGemma: protecting users with state-of-the-art safety classifiers
It is designed to detect and mitigate harmful content in AI models inputs and outputs. ShieldGemma targets four key areas of harm such as, “Hate speech”, “Harassment”, “Sexually explicit content”, and “ Dangerous content”. These open classifiers complement our existing suite of safety classifiers in the Responsible AI Toolkit.
- SOTA performance: built on top of Gemma 2, HieldGemma is the industry-leading safety classifier.
- Flexible sizes: ShieldGemma offers various model sizes to meet diverse needs. The 2B model is ideal for online classification tasks, while the 9B and 27B versions provide higher performance.
- Open and collaborative: the open nature of ShieldGemma encourages transparency and collaboration within the AI community, contributing to the future of ML industry security.
Gemma Scope: illuminating AI decision-making with open sparse autoencoders
Gemma Scope offers researchers and developers unprecedented transparency into the decision-making processes of our Gemma 2 models. Acting like a powerful microscope, Gemma Scope uses sparse autoencoders (SAEs) to zoom in on specific points within the model and make its inner workings more interpretable. Gemma 2 identifies patterns, processes information, and ultimately makes predictions. Here is what makes Gemaa Scope groundbreaking:
- Open SAEs: over 400 freely available SAEs covering all layers of Gemma 2 2B and 9B.
- Interactive demos: explore SAE features and analyze model behavior without writing code on NeuroPedia.
- Easy-to-use repository: code and examples for interfacing with SAEs and Gemma 2.
Therefore these releases represent the ongoing commitment of Google to providing the AI community with the tools and resources needed to build a future where AI benefits everyone.