OTBD: Overfitting tendency assessment based on benchmark deviation for dependable model training

In the context of building dependable intelligent software, the controllable training process is a critical component. Overfitting poses a fundamental risk that directly undermines both the generalization capability and behavioral predictability of models. While early stopping mechanisms help maintain consistent model performance in real-world scenarios through real-time monitoring and intervention, existing methods primarily rely on validation loss or performance stagnation as stopping criteria. These approaches exhibit significant passivity and external dependency, as they struggle to interpret or predict overfitting trend from the perspective of the model's internal parameters and optimization dynamics. To mitigate these limitations, this study proposes an overfitting tendency assessment method based on baseline deviation, which integrates the model's internal dynamics with its external performance signals to enable proactive overfitting risk assessment. This method introduces a Deviation Index to characterize structural and temporal deviation changes during training. Leveraging multi-dimensional features, a hierarchical early-stopping decision mechanism has been constructed. A modular partitioning strategy driven by inter-layer similarity is also incorporated to reduce monitoring complexity. The study demonstrates that the proposed method can identify overfitting trends earlier than conventional approaches. While maintaining stable validation performance, it reduces redundant training epochs by least 11.2%. The proposed method also exhibits robust generalizability and transferability in high-noise and complex scenarios.