On 31st July 2024, Google introduced significant updates to its Gemma AI series, aiming to enhance artificial intelligence safety, performance, and transparency. These updates include releasing three new models: Gemma 2 2B, ShieldGemma, and Gemma Scope. These additions build on the foundation established by the original Gemma 2 series launched in May 2024.
We’re welcoming a new 2 billion parameter model to the Gemma 2 family. 🛠️
It offers best-in-class performance for its size and can run efficiently on a wide range of hardware.
Gemma 2 2B: High Performance with a Small Footprint
The new Gemma 2 2B model stands out for its exceptional performance relative to its size. Despite being a 2 billion parameter model, it has outperformed larger models such as GPT-3.5 in real-world conversational tests. This model is designed to run efficiently across various hardware setups, from edge devices to cloud environments. Optimisations with NVIDIA TensorRT-LLM and integration with popular frameworks like Keras and Hugging Face enhance its flexibility.
DeepMind's new Gemma 2 2B model outperforms GPT-3.5 and Llama 2 70B chat on Chatbot Arena! Insane results attained through distilling larger models.
I already took this small model for a spin on Google AI Studio. Outputs on several preliminary tests look very interesting.… pic.twitter.com/rC7GgYKFfM
“Gemma 2 2B offers unparalleled performance for its size, making it ideal for diverse applications from data centres to local devices,” said a Google spokesperson. The model’s open-access nature facilitates easier experimentation and development, making it accessible even on free-tier GPUs in Google Colab.
ShieldGemma: Enhancing AI Safety
ShieldGemma is a new suite of safety classifiers designed to filter harmful content in AI outputs. This tool targets hate speech, harassment, sexually explicit content, and other dangerous material. ShieldGemma aims to support developers in deploying AI responsibly by providing a robust layer of content moderation.
Google presents ShieldGemma: Generative AI Content Moderation Based on Gemma
– Opensources Gemma2-based content moderation models – Outperform Llama Guard (+10.8% AU-PRC on public benchmarks) and WildCard (+4.3%)
Rebecca Weiss, Executive Director of ML Commons, commented, “As AI matures, the entire industry must invest in developing high-performance safety evaluators. We’re glad to see Google making this investment.”
Gemma Scope: Illuminating AI Decision-Making
Gemma Scope introduces over 400 sparse autoencoders (SAEs) to provide unprecedented insight into the decision-making processes of Gemma 2 models. These tools allow researchers to explore the complex inner workings of the AI, making it easier to understand how the models identify patterns and make predictions.
“Gemma Scope is designed to advance the field of mechanistic interpretability, enabling researchers to delve deeper into model behaviour and enhance overall AI transparency,” said a Google representative.
Conclusion
The latest advancements in Google’s Gemma AI series reflect a commitment to improving AI safety, performance, and transparency. With the introduction of Gemma 2 2B, ShieldGemma, and Gemma Scope, Google aims to provide developers and researchers with powerful tools to create safer and more interpretable AI systems. These efforts are timely, coinciding with broader industry discussions on the responsible deployment of AI technologies.
