All about the revised Green India Mission to increase forest cover, address climate change
International Yoga Day: Why is it hard to pinpoint the origin of yoga?
QA::Why Madhya Pradesh CM's wish to count snakes and rear king cobras is unfeasible
What new Registration Bill says, why it was introduced
After OpenAI’s text-to-video model Sora dominated news feeds last week, Google has now introduced Gemma, its latest open artificial intelligence (AI) offering.
Here is what you need to know about Gemma.
Gemma is a family of lightweight state-of-the-art open models that has been built using the same research and technology used in Gemini models by Google DeepMind, and other teams across Google. Google said that its name is derived from the Latin word ‘gemma,’ which translates to precious stone.
Gemma is being offered in two model sizes, Gemma 2B and Gemma 7B, which have been released with pre-trained and instruction-tuned variants. Along with Gemma, Google has also released a new Responsible Generative AI toolkit that provides guidance and essential tools for creating safer AI applications with Gemma.
For developers, Google is offering toolchains for inference and supervised fine-tuning (SFT) across major frameworks such as JAX, PyTorch, and TensorFlow through native Keras 3.0. The model comes with ready-to-use Colab and Kaggle notebooks along with integration into popular tools such as Hugging Face, NVIDIA, NeMo, MaxText, and TensorRT-LLM.
These integrations make it accessible for almost any developer to get started with Gemma. According to Google’s official statement on Gemma’s launch, the company envisions democratising AI models.
Google has said that the pre-trained and instruction-tuned Gemma models can run on laptops, workstations, or even on Google Cloud with easy deployment on Vertex AI and Google Kubernetes Engine (GKE). Gemma’s optimisation for multiple AI hardware platforms ensures industry-leading performance including NVIDIA GPUs and Google Cloud TPUs.
Google has said that Gemma shares some key technical and infrastructure components with Gemini, which is its most capable AI model to date. And, because of its underlying technology, both Gemma 2B and Gemma 7B are capable of achieving the ‘best-in-class performance’ for their sizes when compared to other open models.
Gemma has reportedly outperformed significantly larger models on key benchmarks while complying with rigorous standards for safe and responsible outputs.
The tech giant shared a list of scores obtained by Gemma 7B when compared to Meta’s Llama 2 7B in areas like reasoning, maths, and code. In reasoning, Gemma scored 55.1, while Llama 2 secured 32.6 in the BBH benchmark. Similarly, in maths, under the GSM8K Gemma scored 46.4 while Llama 2 scored 14.6. The model also overtook Llama 2 when it came to complex problem-solving in maths, with Gemma scoring 24.3 and Llama 2 scoring 2.5 in the MATH 4-shot benchmark. In terms of Python code generation, Gemma scored 32.3 and Llama 2 scored 12.8.
Google claims that Gemma has been designed in compliance with its AI principles. The tech giant has said that to make Gemma pre-trained models safe and reliable, they used automated techniques to filter certain personal information and sensitive data from its training sets.
Google has also fine-tuned Gemma’s models with human feedback to promote responsible behaviours, and conducted thorough evaluations, including manual and automated testing to minimise risk.
Besides this, Google is also providing a toolkit alongside Gemma to help developers prioritise safety in AI applications. This toolkit comes with methods for building safety classifiers, debugging tools, and guidance based on Google’s experience in developing large language models.
With AI growing at a rapid pace, the clamour around its responsible use, and for greater regulation is also increasing. This is why tech corporations investing billions of dollars on AI are also constantly talking about its principles and safety practices.
“While we are optimistic about the potential of AI, we recognise that advanced technologies can raise important challenges that must be addressed clearly, thoughtfully, and affirmatively. These AI Principles describe our commitment to developing technology responsibly and work to establish specific application areas we will not pursue,” Google said on its official website.
According to Google, the objectives for AI applications include the following: the need to be socially beneficial, avoid creating or reinforcing unfair bias, be built and tested for safety, be accountable to people, incorporate privacy design principles, uphold high standards of scientific excellence, and most importantly, being made available for uses that accord with these principles.
Apart from these objectives, Google has also listed the application areas in which it will not be designing or deploying AI. The company said that it will not be designing or deploying AI in technologies that cause or are likely to cause overall harm. AI with the potential for significant harm will be approached cautiously, weighing benefits against risks with safety measures in place. Weapons or tools primarily designed to harm individuals will not be pursued. Surveillance technology that breaches global norms will not be developed. AI with an objective of conflicting with international law and human rights will not be pursued.
The company has said that the list will adapt as it gains more insights into AI applications.
