Google has expanded its Gemma family of open language models with the release of Gemma 2 2B, a compact 2-billion-parameter model designed for efficient deployment on devices and in resource-constrained environments. Alongside the model, Google introduced ShieldGemma, a safety classifier aimed at filtering harmful content, and Gemma Scope, a suite of interpretability tools to help researchers understand model behavior.
Gemma 2 2B is the smallest addition to the Gemma 2 series, which already includes models with 9 billion and 27 billion parameters. Despite its size, Google claims the model delivers strong performance on benchmarks such as MMLU and HellaSwag, making it suitable for tasks like text generation and summarization on edge devices.
ShieldGemma is a set of safety classifiers built on the Gemma 2 architecture, designed to detect hate speech, harassment, and sexually explicit content. It is open-source and can be integrated into developer workflows to moderate inputs and outputs.
Gemma Scope provides tools for analyzing the internal representations of Gemma models, offering insights into how the models process data and enabling researchers to study model transparency and bias.
All three releases are available on Hugging Face and through Google's AI platform. Google emphasizes that these tools are part of its commitment to responsible AI development, providing the community with both powerful models and mechanisms to ensure their safe use.