Google has announced the release of a large language model called Gemma, built using technologies applied in the chatbot model Gemini, which aims to compete with ChatGPT. The model is available in four variants, encompassing 2 and 7 billion parameters, in both base and dialog-optimized formats. The 2 billion parameter variants are suitable for use in consumer applications and can be processed with just a CPU. The 7 billion parameter variants require more powerful hardware, including a GPU or TPU.
Among the applications of the Gemma model are the creation of dialog systems and virtual assistants, text generation, forming responses to questions asked in natural language, summarizing and abstracting content, explaining concepts and terminology, correcting text errors, and assisting in language learning. It supports the creation of various types of text data, including poetry, code in programming languages, rewriting works in different words, and generating letters based on templates. The model is relatively compact, allowing it to be used on hardware with limited resources, such as standard laptops and PCs.
The license for the model permits free use and distribution not only in research and personal projects but also in commercial products. It also allows for the creation and publication of modified versions of the model. However, the usage conditions prohibit the application of the model for malicious activities and recommend using the latest version of Gemma in products whenever possible.
Support for working with the Gemma models has already been added to the Transformers toolkit and the Responsible Generative AI Toolkit. The Keras framework can be used to optimize the model, along with backends for TensorFlow, JAX, and PyTorch. Additionally, there is support for using Gemma with the MaxText, NVIDIA NeMo, and TensorRT-LLM frameworks.
The context size considered by the Gemma model is 8,000 tokens (the number of tokens the model can process and remember when generating text). In comparison, the context size of the Gemini and GPT-4 models is 32,000 tokens, while the GPT-4 Turbo model supports 128,000 tokens. The model supports only the English language. In terms of performance, the Gemma-7B model is slightly less capable than the LLama 2 70B Chat and slightly ahead of the DeciLM-7B, PHI-2 (2.7B), and Mistral-7B models. Compared to Google, the Gemma-7B model is slightly ahead of LLama 2 7B/13B and Mistral-7B.
Source: opennet.ru
