Gemma 2 2B AI Model Released by Google DeepMind, Said to Outperform GPT 3.5 Models in Benchmark

Gemma 2 2B synthetic intelligence (AI) mannequin was launched by Google DeepMind on Thursday. It has turn into the newest addition to the Gemma 2 household of AI fashions and joins Gemma 2 27B and 9B fashions. Despite the light-weight dimension, the corporate claims that it outperformed GPT-3.5 fashions on the LMSYS Chatbot Arena benchmark. Alongside, the tech big additionally launched DefendGemma, a collection of classifier fashions to filter the enter and output of Gemma 2, and Gemma Scope, a analysis software that provides insights into how the AI mannequin capabilities.

Gemma 2 2B AI Model’s Features

In a weblog post on Google for Developers, the corporate introduced Gemma 2 2B, which has turn into the smallest language mannequin of the household. Introduced as an on-device AI mannequin, the put up highlighted that regardless of its small parameter dimension, the output is considerably larger than its weight class because it was distilled from bigger fashions. However, the tech big didn’t reveal which AI fashions had been used for its coaching.

Google additionally claimed that the Gemma 2 2B AI mannequin outperformed the GPT-3.5 fashions on the massive mannequin techniques organisation (LMSYS) Chatbot Arena Elo rating. The AI mannequin is claimed to have acquired a rating of 1126, whereas Mixtral 8x7b Instruct v0.1 mannequin scored 1114 and GPT-3.5 scored 1106.

The AI mannequin has additionally been optimised to run on a variety of {hardware}. For edge units and cloud-based deployment, it has been fine-tuned for Vertex AI and Google Kubernetes Engine (GKE). It can be optimised for Nvidia TensorRT-LLM library and has been made accessible as an Nvidia NIM. Further, Gemma 2 2B additionally integrates with Keras, JAX, Hugging Face, and different main platforms.

Since it’s an open-source AI mannequin, the open weights will be downloaded from Google’s Hugging Face listing, Kaggle, or Vertex AI Model Garden. It may also be tried out on the Google AI Studio.

DefendGemma
Photo Credit: Google

 

Apart from Gemma 2, DefendGemma, a collection of security classifiers that may detect and take away dangerous content material in each the enter and output of the AI mannequin, was additionally launched. Google mentioned the system will give attention to hate speech, harassment, sexually express content material, and harmful content material.

Finally, Gemma Scope, a analysis software for academicians and builders was additionally launched. The system makes use of sparse autoencoders (SAEs) to pinpoint particular elements throughout the mannequin to focus on how the decision-making course of works and the way the structure operates.