Announced this year during Google I/O 2024New open source model Gemma 2, and recently officially provided it to developers and researchersCorresponding to 270 billion parameter versionsLater, Google announced the launch ofGemma 2 2B, ShieldGemma and Gemma Scope includedOpen source models corresponding to different usage requirements.
Among them, Gemma 2 2B is a lightweight model that only operates with 20 billion sets of parameters. It can perform content generation or analysis functions on edge devices such as laptops, and has been recognized by certain research institutions and commercial companies. It is currently available through the Google Vertex AI platform, AI Studio services, or through the data modeling and data analysis competition platform Kaggle.
ShieldGemma focuses on identifying problematic content such as hate speech, harassment, and sexually explicit content. It is built on Gemma 2 and is mainly used to filter content generated by automatically generated artificial intelligence. It can also filter out user attempts to issue problematic instructions.
As for Gemma Scope, it allows developers to more easily understand the operating logic of the Gemma 2 model, and then fine-tune it for specific operational results, making Gemma 2 available in a sufficiently transparent form.
Comparing the usage model of Gemini, Google previously stated that the Gemma model is mainly provided to the outside world in an open source form, which is essentially similar to the Llama model used by Meta for open source. Gemini is mainly used for Google's own products, and is also provided for developers to import and use, but the source code is not made public.
