Abstract: The reverse engineering of gate-level netlists is a detection method to ensure the trust worthiness of integrated circuit (IC) products. With the development trend of AI + IC, it has gradually evolved into a task of identifying basic components in netlists based on graph neural networks (GNN). Currently, one of the main challenges facing this task is the inability to generalize to scenarios where the circuit functionality remains consistent but the netlist structure changes. However, there is no publicly available datasets that consist of functionally-identical but structural-different netlists, which adds obstacles to the research on this issue. To address this, 10 datasets with functionally-identical but structural-different netlists have been researched and released to promote in-depth studies on the task of identifying basic components in netlists. Specifically, different standard cell libraries are first utilizedto alter the netlist structures, and 5 datasets are respectively created to verify the accuracy of GNN model, the scalability across different standard cell libraries, and the generalization ability across different standard cell libraries. Secondly, the built-in commands of the synthesis tool areused to alter the netlist structures, and 5 datasets arecreated to validatethe GNN model's generalization capability with in the same standard cell library. Moreover, netlists typically cannot be used as direct inputs to neural network models. To address this, a tool called PyNetlist is developed to automate the conversion of netlists into graph data formats. Finally, the performance of some classic GNN models on all 10 datasets is presented. The experimental results confirm the real existence of the aforementioned issues and also highlight the foundational supporting role of the datasets in addressing these problems. PyNetlist and all 10 datasets will be open-sourced on the journal's website.