GoogleThe company recently released CodeGemma Advanced Large Language Models (LLMs), specifically designed for code generation, understanding and tracing instructions, etc.The development goal is to make it easier for developers around the world to obtain high-qualityCoding assistance tools.
CodeGemma is launched as an open access collection of Gemma models, optimized and fine-tuned for code development scenarios. The suite includes three different models:
-
2 billion parameter base model: optimizes completion and code generation, focusing on providing fast and efficient solutions for code while paying attention to latency and privacy.
-
7 billion parameter base model: Integrates code completion and natural language processing, enhancing its usefulness in completing code and understanding language generation tasks.
-
7 billion parameter model for guided tracing: Users engage in conversations about code, programming, and mathematical reasoning, and are a valuable resource for developers seeking guidance or annotations.
The above 3 models of CodeGemma utilize pre-trained Gemma checkpoints, and are also trained on another 500 billion word blocks in English, mathematics, and various coding languages. Their extraordinary capabilities in logical and mathematical reasoning set a new benchmark for code generation and completion.
The 7 billion parameter model performs well in various programming languages including Python, Java, JavaScript, and C++. This capability is well demonstrated by the excellent results achieved in the HumanEval and MultiPL-E benchmarks. In addition, the model ranked top in the GSM8K evaluation, which also proves its versatility and effectiveness.
