Google has released the LLM Gemma 2 2B model, a small-sized model designed to run directly on devices. Boasting capabilities surpassing GPT-3.5, it is considered the most efficient model in a similar size range.
Trained on a dataset of 2 trillion tokens, comprising web data, code, and mathematical information, this model’s dataset is significantly smaller than larger models. The test results, such as MMLU at 51.3, are considerably lower compared to larger models, or the HumanEval code-writing test at 17.7. However, performance in the Chatbot Arena test, conducted with real users, yielded excellent scores, outperforming GPT-3.5, ChatGPT, Mixtral 8x7B, and Llama 2 70B.
With this small-sized model, we can run models anywhere, including on the NVIDIA T4 chip provided for free with Google Colab services. In addition to the main Gemma 2 2B model, Google has also released the ShieldGemma model for filtering dangerous content, along with Gemma Scope, a tool to showcase the internal workings of Gemma 2, revealing which words the model focuses on to generate responses.
TLDR: Google introduces the LLM Gemma 2 2B model for direct device usage, with impressive performance metrics, including surpassing larger models in real user tests. Additional models like ShieldGemma and Gemma Scope provide enhanced functionality for content filtering and internal model insight.
Leave a Comment