Google releases open, commercially available, and lightweight large-scale language model ``Gemma''

Google released the open source large-scale language model ' Gemma ' on February 22, 2024. Gemma is characterized by being lighter than the multimodal AI Gemini, and can be used commercially.

Gemma - The most advanced lightweight open model family from Google. | Google AI for Developers

https://ai.google.dev/gemma?hl=ja

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 | Google Cloud Official 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/

There are two types of Gemma: 'Gemma 2B' with 2 billion parameters and 'Gemma 7B' with 7 billion parameters, and each has a 'pre-trained model' and an 'instruction-tuned model' tailored for specific tasks. ' is available. Additionally, Gemma's terms of use permit commercial use and distribution for all organizations, regardless of size.

Gemma supports 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 used on your laptop, workstation, or Can be run on Google Cloud.

Google also says that it has worked with NVIDIA to optimize Gemma for Google Cloud TPUs and NVIDIA GPUs. According to NVIDIA , Gemma will be added to the selectable models of Chat with RTX, a chatbot AI that can operate on NVIDIA's GeForce RTX GPU.

NVIDIA releases free chatbot AI 'Chat With RTX', can operate locally on PC equipped with GeForce RTX GPU - GIGAZINE

In addition, Gemma can be deployed on Google Kubernetes Engine (GKE) and has also been added to Vertex AI Model Garden, a platform that aggregates pre-built machine learning models. This allows developers to easily transform custom-tuned models into scalable endpoints that can power AI applications of any size.

Google compares the performance score of Gemma 7B (green) with LLaMA 2's 7B model (blue) and 13B model (red), and Mistral's 7B model (yellow), and summarizes it in the graph below. In Question Answering and Reasoning, Gemma shows almost the same performance as other models, and in Math/Science and Coding, Gemma outperforms other models. I'm hitting the score.

Below is a table comparing the scores of each benchmark test of Gemma's 2B model and 7B model with LLaMa 2's 7B model and 13B model, and Mistral's 7B model. Looking at the average score, Gemma 7B outperforms other models.

Google is also releasing a Responsible Generative AI Toolkit at the same time as Gemma. Google promotes ``responsible AI development'' based on three points: ``responsible design,'' ``robust and transparent evaluation,'' and ``supporting responsible development.'' 'As part of making Gemma's pre-trained models secure and reliable, we used automated techniques to exclude certain personal information and other sensitive data from the training set.' said.