AI Ecosystem Updates | Issue# 6 [September 02, 2024]
What?
Gemma 2, launched by Google in June 2024, are best-in-class open models that were made available in two model sizes: 27 billion (27B) and 9 billion (9B) parameters. The 27B model has quickly gained recognition for its exceptional performance, ranking among the top models on the LMSYS Chatbot Arena leaderboard. The achievement is notable as it surpasses models that are significantly larger in size in real conversations, demonstrating the model’s efficiency and effectiveness in real-world applications. Feel free to learn more about Gemma 2 in this coverage of ours. Beyond just performance, Gemma 2 emphasizes responsible AI by integrating safety, accessibility, and transparency at its core. Google recently announced three new additions to the Gemma 2 family to support the commitment towards responsible AI. These are Gemma 2 2B, ShieldGemma, and Gemma Scope. Let us learn more about them.
Gemma 2 2B – Lightweight and Highly Performant
The Gemma 2 family has expanded with the introduction of the lightweight Gemma 2 2B model. With 2 billion parameters, the new model is designed to offer a powerful combination of performance and efficiency. Despite its smaller size, the Gemma 2 2B model outperforms larger models like GPT-3.5 in conversational tasks, as evident from its high ranking on the Chatbot Arena. This is attributed to the fact that the model is able to learn from larger models like Gemma 2 9B through knowledge distillation.
The 2B model is optimized for deployment across a wide range of hardware, from edge devices and laptops to cloud-based systems, such as Vertex AI and Google Kubernetes Engine (GKE). To further enhance its speed, it is optimized with the NVIDIA TensorRT-LLM library and is available as an NVIDIA NIM. This optimization targets various deployments, including data centers, cloud, local workstations, PCs, and edge devices. See more here. Additionally, it integrates seamlessly with various development tools, including Keras, JAX, Hugging Face, NVIDIA NeMo, Ollama, Gemma.cpp, ensuring broad accessibility for developers.
Gemma 2 2B has been made accessible under the commercially viable Gemma terms for use across academic research and commercial applications. The lightweight nature of the model allows for easy experimentation using the free tier of T4 GPUs and TPUs in Google Colab. Additionally, one can try the model out using Google AI Studio. The model’s weights are available across Kaggle, Hugging Face, Vertex AI Model Garden.
ShieldGemma: Enhancing AI Safety
ShieldGemma is introduced as a robust solution for enhancing the safety of AI outputs. This suite of classifiers is specifically designed to filter harmful content in AI models, focusing on critical areas, such as hate speech, harassment, sexually explicit material, and dangerous content. The open classifiers enhance the current safety tools in Google’s Responsible AI Toolkit, which features methods for creating custom classifiers with minimal data and includes ready-to-use Google Cloud classifiers accessible through API. Google vouches for ShieldGemma’s state-of-the-art performance and claims that the classifiers are industry-leading safety classifiers. ShieldGemma builds upon the foundation of Gemma 2, offering different model sizes tailored for various applications – the smaller, 2B model for real-time, online tasks and the larger 9B and 27B models for more demanding offline tasks. The open and collaborative nature of ShieldGemma encourages transparency and collective progress within the AI community, contributing to the development of industry safety standards in ML. More here.
Gemma Scope: Transparency in AI Decision-Making
Gemma Scope is a groundbreaking tool designed to bring transparency to the AI decision-making process. It uses Sparse Autoencoders (SAEs) to dissect and analyze the inner workings of the Gemma 2 models. By expanding complex information into more understandable and interpretable forms, Gemma Scope allows researchers and developers to gain insights into how AI models identify patterns, process data, and make predictions. The tool is particularly valuable for researchers and developers looking to build AI systems that are not only powerful but also understandable, accountable, and reliable. Gemma Scope provides over 400 SAEs covering all of the layers of the Gemma 2 2B and 9B models that are openly available, along with interactive demos on Neuronpedia, and easy-to-use code examples built around Gemma 2 and SAEs, facilitating extensive experimentation and use. More about Gemma Scope in this DeepMind blog, technical report, and developer documentation.
Google’s Commitment to Responsible AI – Thoughts
Building a Safer AI Ecosystem
The developments in Gemma 2 highlight Google’s commitment to building a safer and more responsible AI ecosystem. By offering models and tools that prioritize safety, transparency, and accessibility, Google aims to foster an environment where AI can be leveraged for the benefit of all users. The open access to these resources invites collaboration and innovation, ensuring that the advancements in AI are aligned with ethical standards and societal needs.
Empowering Researchers and Developers
With the introduction of Gemma 2 2B, ShieldGemma, and Gemma Scope, developers and researchers are equipped with the tools necessary to create safer, more transparent, and efficient AI applications. Whether it’s through the use of high-performance models that can be deployed on various hardware or through tools that enhance the understanding and safety of AI systems, these advancements provide the foundation for the next generation of responsible AI development.
If you are interested to learn more, feel free to check out this blog by Google.
Thank you for reading through! I genuinely hope you found the content useful. Feel free to reach out to us at [email protected] and share your feedback and thoughts to help us make it better for you next time.
Acronyms used in the blog that have not been defined earlier: (a) Artificial Intelligence (AI), (b) Billion (B), (c) Personal Computer (PC), (d) Graphics Processing Unit (GPU), and (e) Tensor Processing Unit (TPU).