In the era of digital justice, the integration of big data analytics into sentencing decisions has emerged as a key direction for enhancing judicial transparency and fairness. This paper proposes a novel sentencing standardization framework based on judicial big data and interpretable machine learning. Focusing on online fraud adjudication documents from the Chinese judiciary, we construct a domain-specific database using a hybrid method of keyword-based pattern matching and association rule analysis to extract structured features such as criminal intent, means, economic loss, and mitigating factors. These features are encoded into machine-readable vectors and fed into a LightGBM-based gradient boosting decision tree (GBDT) model to predict sentencing outcomes. Extensive experiments using real-world fraud cases demonstrate the model’s high predictive performance, with R² scores reaching 0.98 and minimal average deviation. A series of visual and statistical evaluations—including boxplots, Taylor diagrams, and regression fits—validate the model’s robustness and its ability to replicate human sentencing logic.