Google introduces new state-of-the-art open models

Design

Gemma is designed with our artificial intelligence principles in the foreground. As a part of the pre -trained Gemma models, we used automated techniques to filter some personal information and other sensitive data from training groups. In addition, we used comprehensive learning of light and reinforcement of human comments (RLHF) to align our models that were seized with responsible behaviors. To understand and reduce the risk profile of GEMMA models, we conducted strong assessments including manual red victory, automatic numerical tests, and typical capabilities assessments of dangerous activities. These assessments are shown in our model card.

We also release AI to a new artificial intelligence group with GEMMA to help developers and researchers give priority to building safe and responsible artificial intelligence applications. Includes a set of tools:

• Safety Classification: We offer a new methodology for building strong safety works with minimal examples.
• Correction of errors: The typical error correction tool helps you to investigate the behavior of GEMMA and address potential issues.
• guidance: You can reach best practices for models based on the Google experience in developing and publishing large language models.

It has been improved through the frameworks, tools and devices

You can adjust the GEMMA models on your own data to adapt to the specific application needs, such as the summary or the generation of retrieval (RAG). Gemma supports a wide range of tools and systems:

• Multiple spears: Bring your favorite Ettra, with reference applications for inference and control via Keras 3.0, original Pytorch, Jax, and Luging Face Transformers.
• Overly compatibility: Gemma models work through common devices, including laptop, desktop, internet, mobile phone and cloud, allowing the possibilities of artificial intelligence widely available.
• Advanced devices ’platforms: We have partnership with NVIDIA to improve GEMMA for NVIDIA GPU, from data center to cloud to local local computers, ensuring the leading performance in industry and integration with advanced technology.
• The optimum for Google Cloud: Vertex Ai provides a wide MLOPS tool collection with a set of tuning options and click using the improvements for inference. Advanced allocation is available with fully managed AI Vertex or GKE tools, including publishing on cost -cost infrastructure via GPU, TPU and CPU of either of the two systems.

Free credits for research and development

Gemma is designed for the open community of developers and researchers who create artificial intelligence. You can start working with GEMMA today with free access in Kagge, free Tier for the Colab Notebooks, and $ 300 in credits for Google Cloud users for the first time. Researchers can also apply for Google Cloud assets of up to $ 500,000 to accelerate their projects.

Start

You can explore more about GEMMA and reach Quickstart evidence on ai.google.dev/gemma.

As we continue to expand the Gemma Model family, we look forward to introducing new variables for various applications. Stay tuned for events and opportunities in the coming weeks to communicate, learn and build with GEMMA.

We are excited to see what you create!