MAG-Bench

In order to make up for the problems that existing test benchmarks are mostly oriented to textual modalities, the inter-class interaction ability evaluation is insufficient, and the model knowledge base is often leaked, we proposed MAgent_Bench, a python code generation benchmark specifically used to evaluate the inter-class interaction function based on MAG graphs. This benchmark not only inherits the advantages of existing code generation model performance evaluation benchmarks, but also emphasizes the unique characteristics of object-oriented programming, the code generation ability under different modalities based on MAG graphs, and the complex interaction between agent models, aiming to provide a more comprehensive and in-depth evaluation platform for code generation tasks.