LLM Generalization Testing Tool for MindSpore

Developed a component-based generalization testing tool for large-scale models on MindSpore, enabling automated mutation and recombination of model structures such as layers, shapes, and parameters under defined constraints.

Built cross-framework validation pipelines by generating equivalent training scripts in PyTorch and conducting comparative analysis to detect training failures and accuracy anomalies. Designed iterative strategies to identify fault-inducing factors and support multi-dimensional issue diagnosis across functionality, performance, and numerical accuracy.