The Knowledge Computing Laboratory of Peking University and the AI team of Sichuan Tianfu Bank officially open sourced CodeShell, a large code model with 7 billion parameters, and provided deployment solutions and IDE plug-ins that match CodeShell. Please view the CodeShell code base for details .
According to the official introduction, CodeShell has 7 billion parameters, was trained with 500 billion Tokens, and the context window length is 8192. CodeShell's original training data is based on the GitHub data, Stack and StarCoder data sets crawled by itself, as well as a small amount of high-quality Chinese and English data.
In authoritative code evaluation benchmarks (HumanEval and MBPP), CodeShell achieves the best performance of the same scale.
The open source model this time is as follows:
- CodeShell Base : CodeShell base model, with powerful code base capabilities.
- CodeShell Chat : CodeShell conversation model, with excellent performance in downstream tasks such as code Q&A and code completion.
- CodeShell Chat 4bit : A 4bit quantified version of the CodeShell conversation model, which consumes less memory and is faster while ensuring model performance.
- CodeShell CPP : CodeShell conversation model CPP version, supports developers to use on personal computers without GPU. Note that the CPP version also supports quantization operations, and users can run CodeShell on a personal computer with a minimum memory of 8G.
The picture below is a screenshot of the VS Code plug-in of the CodeShell code assistant:
IDEA plug-in:
