How to Use Gemma, Google's New State-of-the-Art Open Models

Google has recently introduced a new generation of open models called Gemma, which are based on the same research and technology used to create the Gemini models. Gemma is a family of lightweight, state-of-the-art open models that are text-to-text, decoder-only large language models, available in English, with open weights, pre-trained variants, and instruction-tuned variants¹. In this blog post, we will explain what Gemma is, why it is important, and how you can use it to build AI applications responsibly.

What is Gemma and why is it important?

Gemma is inspired by Gemini, Google's largest and most capable AI model widely available today. Gemini is a massive, multi-modal, multi-task model that can generate natural language, images, audio, and video from any input modality². Gemini is also instruction-tuned, which means it can perform various tasks by following natural language instructions. However, Gemini is not open to the public and requires specialized hardware and infrastructure to run.

Gemma, on the other hand, is designed to be accessible and useful for everyone. Gemma models share technical and infrastructure components with Gemini, but are much smaller and faster, making them suitable for running on laptops, workstations, or cloud platforms¹. Gemma models also inherit the instruction-tuning capability of Gemini, which enables them to perform a wide range of tasks across domains and modalities¹.

Gemma is important because it represents a new milestone for the field of AI, as it demonstrates that high-quality and versatile models can be achieved at a smaller scale and with lower computational costs. Gemma also contributes to the open community, as it allows developers and researchers to leverage Google's cutting-edge technology and research to create AI applications for various purposes and domains. Moreover, Gemma is designed with Google's AI principles at the forefront, which means it adheres to rigorous standards for safe and responsible outputs¹.

How to use Gemma to build AI applications responsibly?

Gemma is available worldwide, starting from February 21, 2024¹. You can find the model weights, tools, and documentation on [Google AI](^5^), [GitHub](^6^), [Kaggle](^9^), and [Vertex AI Model Garden]. Google provides two sizes of models: Gemma 2B and Gemma 7B, each with pre-trained and instruction-tuned variants¹. You can choose the model size and variant that best suits your needs and resources.

To use Gemma, you will need to provide it with natural language instructions and inputs, and it will generate natural language outputs. For example, you can ask Gemma to summarize a text, translate a sentence, write a poem, or answer a question. You can also combine multiple instructions to perform more complex tasks, such as writing a blog post, creating a presentation, or generating a report. You can use Gemma's outputs as they are, or edit them for further improvement.

However, using Gemma also comes with some responsibilities and challenges. As with any AI model, Gemma is imperfect and may produce inaccurate, incomplete, or inappropriate outputs. Therefore, you should always verify and validate Gemma's outputs before using them for any purpose. You should also be aware of the potential risks and harms that Gemma's outputs may cause to yourself, others, or the environment, and take appropriate measures to prevent or mitigate them.

To help you use Gemma responsibly, Google provides a [Responsible Generative AI Toolkit], which offers guidance and essential tools for creating safer AI applications with Gemma. The toolkit includes:

- A code of conduct that outlines the ethical principles and best practices for using Gemma.

- A data card that summarizes the key information and characteristics of Gemma models, such as their capabilities, limitations, biases, and intended uses.

- A safety testing tool that allows you to evaluate Gemma's outputs on various dimensions of safety, such as toxicity, fairness, privacy, and robustness.

- A feedback mechanism that enables you to report any issues or concerns you encounter while using Gemma.

By using the toolkit, you can ensure that you are using Gemma in a way that aligns with Google's AI principles and respects the rights and interests of all stakeholders.

Conclusion

Gemma is a new family of open models from Google that are based on the same research and technology used to create the Gemini models. Gemma models are lightweight, state-of-the-art, and instruction-tuned, which makes them versatile and powerful tools for building AI applications. However, using Gemma also requires responsibility and caution, as it may produce outputs that are not accurate, complete, or appropriate. Therefore, you should always check and validate Gemma's outputs, and use the Responsible Generative AI Toolkit to create safer AI applications with Gemma.