Google releases Gemma, an open, commercially available, lightweight, large-scale language model
Google released the open-source large-scale language model ' Gemma ' on February 22, 2024. Gemma is lighter than the multimodal AI Gemini and can also be used commercially.
Gemma - Google's cutting-edge, lightweight, open model family. | Google AI for Developers
Gemma: Open Models Based on Gemini Research and Technology
(PDF file) https://storage.googleapis.com/deepmind-media/gemma/gemma-report.pdf
Gemma now available on Google Cloud | Official Google Cloud Blog
https://cloud.google.com/blog/ja/products/ai-machine-learning/gemma-model-available-in-vertex-ai-and-via-gke/
Gemma: Google introduces new state-of-the-art open models
https://blog.google/technology/developers/gemma-open-models/
Gemma is available in two versions: 'Gemma 2B' with 2 billion parameters and 'Gemma 7B' with 7 billion parameters. Each version offers both pre-trained models and instruction-tuned models tailored to specific tasks. Gemma's terms of use also allow commercial use and distribution by all organizations, regardless of size.
Gemma supports the tools used by developers at Google Cloud, including Google Colab and Kaggle Notebook , as well as the JAX , PyTorch , Keras 3.0 , and Hugging Face Transformers frameworks, and can be run on a laptop, workstation, or on Google Cloud.
Google also said it has worked with NVIDIA to optimize Gemma for Google Cloud TPUs and NVIDIA GPUs, and that Gemma will be added as an available model for its chatbot AI 'Chat with RTX,' which runs on NVIDIA's GeForce RTX GPUs.
NVIDIA releases free chatbot AI 'Chat With RTX,' which can run locally on PCs equipped with GeForce RTX GPUs - GIGAZINE
Additionally, Gemma can be deployed on Google Kubernetes Engine (GKE) and has been added to the Vertex AI Model Garden, a platform of pre-built machine learning models that enables developers to easily transform their custom-tuned models into scalable endpoints that can power AI applications at any scale.
Google compares the performance scores of Gemma 7B (green) with the LLaMA 2 7B model (blue) and 13B model (red), and Mistral 7B model (yellow) in the graph below. In Question Answering and Reasoning, Gemma performs roughly on par with the other models, but in Math/Science and Coding, Gemma outperforms the other models.
Below is a table comparing the benchmark test scores of Gemma's 2B and 7B models with the LLaMa 2 7B and 13B models and the Mistral 7B model. Looking at the average score, Gemma 7B outperforms the other models.
At the same time as Gemma, Google also released the ' Responsible Generative AI Toolkit .' The toolkit promotes responsible AI development, consisting of three key points: 'responsible design,' 'robust and transparent evaluation,' and 'supporting responsible development.' The company stated, 'Gemma is designed with AI principles as a top priority. As part of making Gemma's pre-trained models safe and reliable, we used automated techniques to remove certain personal information and other sensitive data from the training set.'
Related Posts:
