Google has announced the publication of Gemma, a large machine learning language model built using technologies used to build the Gemini chatbot model, which attempts to compete with ChatGPT. The model is available in four variants, covering 2 and 7 billion parameters, in basic and dialog-optimized views. Options with 2 billion parameters are suitable for use in consumer applications and have enough CPU to process them. Options with 7 billion parameters require more powerful hardware and a GPU or TPU.
Among the areas of application of the Gemma model are the creation of dialogue systems and virtual assistants, text generation, generation of answers to questions asked in natural language, summary and generalization of content, explanation of the essence of concepts and terms, correction of errors in the text, assistance in learning languages. It supports the creation of various types of text data, including poetry, code in programming languages, rewriting works in other words, and generating letters using a template. Moreover, the model has a relatively small size, allowing it to be used on your own equipment with limited resources, for example, on ordinary laptops and PCs.
The model license allows free use and distribution not only in research and personal projects, but also in commercial products. The creation and publication of modified versions of the model is also allowed. At the same time, the terms of use prohibit the use of the model to perform malicious actions and require, whenever possible, to use the latest version of Gemma in your products.
Support for working with Gemma models has already been added to the Transformers toolkit and the Responsible Generative AI Toolkit. To optimize the model, you can use the Keras framework and backends for TensorFlow, JAX and PyTorch. It is also possible to use Gemma with the MaxText, NVIDIA NeMo and TensorRT-LLM frameworks.
The size of the context taken into account by the Gemma model is 8 thousand tokens (the number of tokens that the model can process and remember when generating text). For comparison, the context size for the Gemini and GPT-4 models is 32 thousand tokens, and for the GPT-4 Turbo model it is 128 thousand. The model only supports English. In terms of performance, the Gemma-7B model is slightly inferior to the LLama 2 70B Chat model and slightly ahead of the DeciLM-7B, PHI-2 (2.7B) and Mistral-7B-v0.1 models. In the Google comparison, the Gemma-7B model is slightly ahead of the LLama 2 7B/13B and Mistral-7B.
Source: opennet.ru