Gemma, a new open source artificial intelligence model
Google announced, through a blog post, the launch of his new family of AI models based on the Gemini chatbot, «Gemma«. This is a machine learning model that is built on the technologies used for Gemini, Google's chatbot model and offers a range of variants ranging from 2 to 7 billion parameters, designed for different applications and hardware requirements.
Gemma aims to provide to developers advanced tools for creating AI applications consciously and among the areas of application that Gemma covers, it is mentioned from the creation of dialogue systems and virtual assistants to the generation of text, answers to questions in natural language, content summaries, text correction and learning support of languages. Additionally, the model allows the manipulation of various types of text data, including poetry, programming code, text rewriting, and letter generation using templates.
And is that A highlight of Gemma is her relatively small size, which facilitates its implementation on hardware with limited resources, such as standard laptops and PCs. In comparisons conducted by Huggingface and Google, the Gemma-7B model has demonstrated solid performance, ranking second after the LLama 2 70B Chat model in Huggingface's comparison. In the Google comparison, Gemma-7B is slightly ahead of the LLama 2 7B/13B and the Mistral-7B.
On the part of ecosystem of tools and frameworks, the new AI chatbot offers integration with a large number of tools commonly used by developers, since it has several important projects that have already integrated support to work with Gemma and among the projects that already have support, the following stand out: Hugging Face, MaxText, NVIDIA NeMo, TensorRT-LLM, Transformers, Responsible Generative AI Toolkit among others.
In addition, Google has released a standalone output engine called gemma.cpp, written in C++, specifically for Gemma, and support for Gemma has been added to the llama.cpp engine. To optimize the model, developers can leverage the Keras framework and backends for TensorFlow, JAX, and PyTorch.
It's important to know that The Gemma model has a size of 8 thousand tokens, which limits the amount of information it can process and remember during text generation (for comparison, models like Gemini and GPT-4 have context sizes of 32 thousand tokens, and GPT-4 Turbo has 128 thousand). Additionally, the Gemma model currently only supports English as a language.
To ensure the highest safety standards, Google used automated techniques to remove personal information from the data training of Gemma models. Additionally, reinforcement learning, guided by human feedback, was used to refine the Gemma variants tailored to the instructions, ensuring they adhere to responsible behavioral patterns.
Google mentions that the constantly evolving nature of AI raises important considerations about security and ethical use, since in the wrong hands, the lack of restrictions on open AI models can create significant risks for society. Google recognizes these challenges and has taken a comprehensive approach to addressing them through rigorous assessments and clear terms of use, the company seeks to ensure that open AI models are used ethically and responsibly, while encouraging innovation and collaboration in the community.
For those interested, you should know that Gemma is available in two configurations, Gemma 2B and Gemma 7B, this open source AI model offers variants pre-trained and tuned by instructions to operate efficiently. In addition, Gemma's license allows free use in research, personal and commercial projects, as well as the creation and distribution of modified versions of the model.
finally if you are interested in knowing more about it, you can check the details in the following link