CodeGemma, an open-source code model family built based on Google DeepMind's Gemma model, has attracted significant attention in the fields of code generation and understanding. This model not only enhances code generation capabilities but also retains powerful natural language understanding abilities, providing strong support for various application scenarios. In this paper, we will delve into the technical details and evaluation results of CodeGemma, as well as its impact on the field of AI.
1 Introduction
The CodeGemma model is built on the Gemma pre-trained model and has achieved a leading position in code completion and generation tasks through further pre-training on a large amount of code data. At the same time, it also retains the powerful natural language understanding capabilities of the Gemma model. CodeGemma consists of a 7B parameter pre-trained model, a 7B parameter instruction fine-tuning model, and a 2B parameter model specifically designed for code completion and generation tasks.
Simple test on Ollama
It was found that Ollama already supports codegemma: https://ollama.com/library/codegemma, defaulting to 4-bit quantization, and also supporting other versions. I chose the largest model that can run on a 16GB graphics card, "codegemma:7b-instruct-q6_K".
Online experience: https://huggingface.co/blog/codegemma#demo