Gemma 2 2B AI Model from Google DeepMind Outperforms GPT 3.5 Models in Benchmark
Gemma 2 2B is an open-source AI model available from Kaggle, Hugging Face, and Vertex AI Model Garden.
Google DeepMind unveiled Gemma 2 2B AI model on Thursday. It joins Gemma 2 27B and 9B AI models as the newest addition. The business says it beat GPT-3.5 models on the LMSYS Chatbot Arena test despite its lightweight size. The tech giant also published ShieldGemma, a suite of classifier models to filter Gemma 2 input and output, and Gemma Scope, a research tool that reveals how the AI model works.
Features
Google introduced Gemma 2 2B, the family’s smallest language model, in a Developers blog post. The article described an on-device AI model distilled from larger models. It produced outputs exceeding its weight class despite its modest parameter set. The tech firm did not divulge its training AI models.
The Gemma 2 2B AI model surpassed the GPT-3.5 models on the LMSYS Chatbot Arena Elo score, according to Google. GPT-3.5 scored 1106, Mixtral 8x7b Instruct v0.1 scored 1114, and AI scored 1126.
The AI model is optimized for various hardware, including Vertex AI and Google Kubernetes Engine, for both edge devices and cloud deployment. It is an Nvidia NIM and optimized for Nvidia TensorRT-LLM library. Gemma 2 2B interfaces with Keras, JAX, Hugging Face, and other important platforms.
You can find the open weights for this open-source AI model on Google’s Hugging Face listing, Kaggle, or Vertex AI Model Garden. You can test it on Google AI Studio.
Gemma 2 introduced ShieldGemma, a suite of safety classifiers that identifies and eliminates dangerous information from AI model input and output. Google claimed the algorithm would target hate speech, harassment, sexually explicit material, and risk.
Finally, Gemma Scope, an academic and developer research tool, was launched. Sparse autoencoders (SAEs) emphasize certain model sections to demonstrate the decision-making process and architecture.
CREDIT: Allneeds, Gadgets360
READ RELATED POSTS >>
No products in the cart.