Smaller, Safer, More Transparent: Advancing Responsible AI with Gemma

In June, we released Gemma 2, our new best-in-class open fashions, in 27 billion (27B) and 9 billion (9B) parameter sizes. Since its debut, the 27B mannequin rapidly turned one of many highest-ranking open fashions on the LMSYS Chatbot Arena leaderboard, even outperforming standard fashions greater than twice its measurement in actual conversations.

However Gemma is about extra than simply efficiency. It is constructed on a basis of accountable AI, prioritizing security and accessibility. To help this dedication, we’re excited to announce three new additions to the Gemma 2 household:

1. Gemma 2 2B – a brand-new model of our standard 2 billion (2B) parameter mannequin, that includes built-in security developments and a strong steadiness of efficiency and effectivity.

2. ShieldGemma – a set of security content material classifier fashions, constructed upon Gemma 2, to filter the enter and outputs of AI fashions and preserve the consumer secure.

3. Gemma Scope – a brand new mannequin interpretability device that provides unparalleled perception into our fashions’ interior workings.

With these additions, researchers and builders can now create safer buyer experiences, acquire unprecedented insights into our fashions, and confidently deploy highly effective AI responsibly, proper on machine, unlocking new prospects for innovation.

Gemma 2 2B: Expertise Subsequent-Gen Efficiency, Now On-Gadget

We’re excited to introduce the Gemma 2 2B model, a extremely anticipated addition to the Gemma 2 household. This light-weight mannequin produces outsized outcomes by studying from bigger fashions by way of distillation. In actual fact, Gemma 2 2B surpasses all GPT-3.5 fashions on the Chatbot Enviornment, demonstrating its distinctive conversational AI skills.

Gemma 2 2B gives:

• Distinctive efficiency: Delivers best-in-class efficiency for its measurement, outperforming different open fashions in its class.

• Versatile and cost-effective deployment: Run Gemma 2 2B effectively on a variety of {hardware}—from edge gadgets and laptops to strong cloud deployments with Vertex AI and Google Kubernetes Engine (GKE). To additional improve its pace, it’s optimized with the NVIDIA TensorRT-LLM library and is on the market as an NVIDIA NIM. This optimization targets varied deployments, together with information facilities, cloud, native workstations, PCs, and edge gadgets — utilizing NVIDIA RTX, NVIDIA GeForce RTX GPUs, or NVIDIA Jetson modules for edge AI. Moreover, Gemma 2 2B seamlessly integrates with Keras, JAX, Hugging Face, NVIDIA NeMo, Ollama, Gemma.cpp, and shortly MediaPipe for streamlined growth.

Beginning in the present day, you may obtain Gemma 2’s mannequin weights from Kaggle, Hugging Face, Vertex AI Model Garden. You can even strive its capabilities in Google AI Studio.

ShieldGemma: Defending Customers with State-of-the-Artwork Security Classifiers

Deploying open fashions responsibly to make sure partaking, secure, and inclusive AI outputs requires vital effort from builders and researchers. To assist builders on this course of, we’re introducing ShieldGemma, a collection of state-of-the-art security classifiers designed to detect and mitigate dangerous content material in AI fashions inputs and outputs. ShieldGemma particularly targets 4 key areas of hurt:

• Sexually express content material

These open classifiers complement our current suite of security classifiers within the Responsible AI Toolkit, which features a methodology to construct classifiers tailor-made to a particular coverage with restricted variety of datapoints, in addition to current Google Cloud off-the-shelf classifiers served by way of API.

This is how ShieldGemma may help you create safer, higher AI purposes:

• SOTA efficiency: Constructed on high of Gemma 2, ShieldGemma are the industry-leading security classifiers.

• Versatile sizes: ShieldGemma gives varied mannequin sizes to fulfill numerous wants. The 2B mannequin is right for on-line classification duties, whereas the 9B and 27B variations present greater efficiency for offline purposes the place latency is much less of a priority. All sizes leverage NVIDIA pace optimizations for environment friendly efficiency throughout {hardware}.

• Open and collaborative: The open nature of ShieldGemma encourages transparency and collaboration throughout the AI group, contributing to the way forward for ML {industry} security requirements.

“As AI continues to mature, your complete {industry} might want to spend money on growing excessive efficiency security evaluators. We’re glad to see Google making this funding, and stay up for their continued involvement in our AI Security Working Group.” ~ Rebecca Weiss, Govt Director, ML Commons

Study extra about ShieldGemma, see full ends in the technical report, and begin constructing safer AI purposes with our complete Responsible Generative AI Toolkit.

Gemma Scope: Illuminating AI Choice-Making with Open Sparse Autoencoders

Gemma Scope gives researchers and builders unprecedented transparency into the decision-making processes of our Gemma 2 fashions. Appearing like a robust microscope, Gemma Scope makes use of sparse autoencoders (SAEs) to zoom in on particular factors throughout the mannequin and make its interior workings extra interpretable.

These SAEs are specialised neural networks that assist us unpack the dense, advanced info processed by Gemma 2, increasing it right into a type that is simpler to research and perceive. By learning these expanded views, researchers can acquire beneficial insights into how Gemma 2 identifies patterns, processes info, and finally makes predictions. With Gemma Scope, we purpose to assist the AI analysis group uncover find out how to construct extra comprehensible, accountable, and dependable AI techniques.

This is what makes Gemma Scope groundbreaking:

• Interactive demos: Discover SAE options and analyze mannequin habits with out writing code on Neuronpedia.

Study extra about Gemma Scope on the Google DeepMind blog, technical report, and developer documentation.

A Future Constructed on Accountable AI

These releases symbolize our ongoing dedication to offering the AI group with the instruments and sources wanted to construct a future the place AI advantages everybody. We consider that open entry, transparency, and collaboration are important for growing secure and useful AI.

Get Began Right this moment:

• Strive Gemma Scope on Neuronpedia and uncover the interior workings of Gemma 2.

Be a part of us on this thrilling journey in the direction of a extra accountable and useful AI future!