Gemma, a new AI model, has been crafted with a strong emphasis on safety and reliability, adhering to AI Principles through automated data filtering, fine-tuning, and reinforcement learning from human feedback. To mitigate risks, the model underwent thorough evaluations, including red-teaming and adversarial testing. Alongside Gemma, a Responsible Generative AI Toolkit is introduced, offering safety classification, debugging tools, and best practice guidance for developers.
Gemma, a new AI model, has been crafted with a strong emphasis on safety and reliability, adhering to AI Principles through automated data filtering, fine-tuning, and reinforcement learning from human feedback. To mitigate risks, the model underwent thorough evaluations, including red-teaming and adversarial testing. Alongside Gemma, a Responsible Generative AI Toolkit is introduced, offering safety classification, debugging tools, and best practice guidance for developers.Gemma is versatile, supporting various frameworks like Keras 3.0, PyTorch, JAX, and Hugging Face Transformers, and is compatible across devices and optimized for NVIDIA GPUs and Google Cloud’s Vertex AI. It allows for fine-tuning on specific datasets and is accessible on multiple platforms, including laptops and IoT devices.
Developers and researchers can start experimenting with Gemma for free through Kaggle, Colab notebooks, and Google Cloud credits, with substantial funding available for qualifying research projects. Resources and guides for Gemma are available at ai.google.dev/gemma, with future updates and community events anticipated.
Read more at Google…