Gemma 2B and 7B: Google presents two open-source AI models
If Google seemed to have fallen behind in the AI race during 2023, the firm has been multiplying its announcements for several weeks now. After launching its new Gemini language model (and its improved version, Gemini 1.5), then rebranding its Google Bard chatbot, also named Gemini, Google has announced the launch of two lighter, open-source models: Gemma 2B and Gemma 7B.
Gemma means "precious stone" in Latin, but it's hard not to see a nod to Gemini. Indeed, Gemma 2B and Gemma 7B have been created "using the same research and technologies used to create the Gemini models," says Google in a blog post. The company also points out that "each size is released with pre-trained variants tailored to the instructions.".
Language models for researchers and developers
Both models, available with a commercial license, are specifically targeted at researchers and developers, enabling them to design their own artificial intelligence applications. Gemma will be accessible via Kaggle (now free of charge), Hugging Face, Nvidia's NeMo, and Google's Vertex AI.
To "accelerate university research," a credit will also be granted to researchers on Google Cloud. This could amount to $500,000. A form has already been put online for those wishing to apply for this credit.
Google takes direct aim at Meta.
At the launch of the Gemini model, Google clearly stated its ambition to compete with OpenAI, presenting a comparison between Gemini and GPT-4. For Gemma, the designated rival is Meta, with its open-source Llama 2 language. In a new comparison table (see below), Google aims to demonstrate that Gemma 7B outperforms Llama 7B and 13B in a number of areas: MMLU*, reading comprehension, Python code generation, etc.
Gemma will also be competing with the French company Mistral AI, which specializes in open-source models (see image one). For the time being, Gemma is primarily designed for use in the English language.
With Gemma, Google aims to compete directly with Llama-2. © Google
* The MMLU, or Massive Multitask Language Understanding, is a comprehensive test evaluating AI models on 57 varied tasks, measuring their multitasking understanding of language in various academic and professional domains.
Responsible AI development
In its press release as well as on its presentation page, Google stresses the importance of responsibility in the development of its new language models. The company specifies that the programs have been "pre-trained on carefully selected, safety-aware data." To encourage developers to adopt an ethical approach, Google has also released a Responsible Generative AI Kit. This kit offers resources for using open models like Gemma, including advice on setting safety standards, adjusting safety parameters, safety classifiers, and a learning interpretation tool to identify and resolve potential problems.
When compiling your application with Gemma, you need to take a holistic approach to responsibility and consider all possible challenges at the application and model levels. This kit introduces risk mitigation techniques, security, privacy, impartiality, and liability issues," says Google in the preamble to its kit.