Programmer AI assistant new competition, Ant Group announced the opening of the source code large model CodeFuse
CTOnews.com September 8 news, at present there are many AI products on the market that can assist in writing code, in today's Bund Conference sub-forum, Ant Group for the first time open source code large model CodeFuse, join this market competition.
According to introduction, this is an ant self-developed code generation exclusive large model, which can provide intelligent suggestions and real-time support according to developer input, help developers automatically generate code, automatically add comments, automatically generate test cases, repair and optimize code, etc., so as to improve R & D efficiency.
Ant Group said: "Whether the user is a beginner or an experienced developer, CodeFuse can greatly improve the efficiency and accuracy of programming, making everyone programmable and innovative." "
CodeFuse is based on ant base large model development. In the recent HumanEval evaluation of code completion, CodeFuse scored 74.4%, exceeding the score of GPT-4 (67%) and exceeding the score of WizardCoder-34B 73.2%. This open source content includes code framework, model, etc., which is now available on relevant platforms. CTOnews.com partners can download models at HuggingFace.
The code framework section includes a self-developed Multi-task fine-tuning (MFT) framework. MFT supports code generation, code translation, test case generation, bug fixing, etc. The MFT fine-tuning framework supports ant self-developed models, as well as code capability fine-tuning for multiple open source large models. At the same time, the framework contains innovative fine-tuning data packaging technology, which claims to increase the fine-tuning speed by about 8 times, and uses multiple deployment optimization technology to accelerate inference by about 2 times.
The model part includes CodeFuse 13B-4K and CodeFuse-CodeLlaMa 34B-MFT. During the model training process, CodeFuse uses program analysis and verification technology to filter high-quality code data and construct a code domain-specific dictionary with an average code length of 10%+.
Ant Group revealed that CodeFuse began internal testing as early as June this year, using a real development environment, which can help developers to program with low thresholds, such as writing H5 Mini games in natural language, rapidly developing Alipay Mini programs, etc. Application scenarios based on CodeFuse include development assistants, IDE plug-ins, data analyzers, etc., covering the main requirements of current R & D work, and have been verified in the internal R & D process of Ant Group.
