Google expands Gemma 2 with new models and tools that prioritize security, accessibility and transparency

Aug. 1 (Portaltic/EP) –

Google has expanded the Gemma 2 language model family with three new versions focused on security and accessibility that seek to facilitate the deployment of responsible artificial intelligence (AI).

Gemma is the family of open source AI models which Google first introduced in February in two sizes: 2 billion parameters (2B) and 7 billion parameters (7B), and which it updated in May at its annual developer conference with the launch of Gemma 2.

This update, meanwhile, is available with 27 billion parameters (27B), a size that offers performance on par with Meta’s Llama 3, which has 70 billion parameters.

Now, Google has expanded the Gemma 2 family with new models and tools that “prioritize security and accessibility”: Gemma 2 2B, ShieldGemma and Gemma Scope, As reported in your Official Blog.

Gem 2 2B It is a reduced size version of the main model, with 2 billion parameterswhich integrates security advances and offers performance and efficiency. According to Google, it has superior conversational capabilities than GPT-3.5 in Chatbot Arena.

ShieldGemmafor its part, focuses on the protection of users and uses for them Security classifiers that detect and mitigate harmful content (hate speech, harassment, sexually explicit content, and dangerous content) in AI models, both in inputs and outputs.

This version is offered in different sizes, being available in 2B, 9B and 27B. The latter two are the ones that have the best performance in ‘offline’ applications that do not depend so much on latency.

Finally, Gemma Scope It is actually a tool that allows Check the decision-making processes of the Gemma 2 family to improve transparency. Identifies patterns, processes information and makes predictions.