Gemma: Introducing new state-of-the-art open models

Feb 21, 2024 [read-time] In a technologically advanced world, where artificial intelligence (AI) models are being developed and used to assist developers and researchers in building responsible AI applications, Google has introduced Gemma, a new generation of open models built from the same research and technology used to create Gemini models. The newly launched models share technical and infrastructure components with Gemini, making them compatible with other open models that developers can utilize as they develop and launch their own applications. Developers are free to work with Gemma today using Google Cloud’s free tier for Colab notebooks and $300 in credit offers for first-time Google Cloud users. Researchers can also apply for credit offers up to a collective $500,000 to accelerate their projects. Gemma models, designed with AI Principles at the forefront, are capable of achieving best-in-class performance on benchmarks compared to other open models, and surpass significantly larger models by performing on key benchmarks more efficiently than those in the Gemini family that share technical and infrastructure components with Gemma. As part of making Gemma pre-trained models safe and reliable, Google used automated techniques such as filters and fine-tuning to remove sensitive data from training sets, and performed robust evaluation using extensive fine-tuining and reinforcement learning to align instructional and dangerous behavior. The Model Card provides insights into the risk profile for Gemma models and suggests best practices for responsible AI applications. Responsible Generative AI Toolkit is also launching alongside Gemma, providing developers and researchers with the necessary tools and systems to prioritize building safe and responsible AI applications. The toolkit includes access to free Colab notebooks, a $300 in credit offer for first-time Google Cloud users, and $500 credit offers for Gemini model owners. Researchers can apply for the $300 credit offer or apply for the collective credit offer of up to $500,000 for fast-track access to AI-related resources, such as data, software, and hardware, on Google Cloud. The toolkit also includes a researcher portal where developers can find information about responsible generative AI models that adhere to industry standards, such as the AI Principles. To help developers and researchers prioritize building safe and responsible AI applications, Google has released an extensive Responsible Generative AI Toolkit for free on Colab notebooks. In conclusion, Gemma is a family of lightweight, state-of-the-art open models developed by Google, which share technical and infrastructure components with Gemini, making them compatible with other open models. The new models are capable of achieving best-in-class performance for their sizes compared to other open models, and surpass significantly larger models on key benchmarks while adhering to AI Principles at the forefront. As part of making Gemma pre-trained models safe and reliable, Google used automated techniques such as filters and fine-tuning to remove sensitive data from training sets. The Model Card provides insights into the risk profile for Gemma models, which suggests best practices for responsible AI applications. The Responsible Generative AI Toolkit is now available on Colab notebooks, providing developers and researchers with easy access to the necessary tools and systems for building safe and responsible AI applications.