Abstract: Reinforced concrete (RC) frame is predominant of China’s seismic-resistant building systems. The credibility of its fragility model directly affects the accuracy of earthquake risk assessment. Although significant advances have been made in seismic fragility modeling of RC frames, the results are mostly based on numerical simulations, lacking systematic integration and validation. This study developed a machine learning-driven cleaning method. Based on a database comprising fragility curves from Chinese literature, the mapping relationship between structural parameters and spectral acceleration S_am(T₁) was firstly established by machine learning models, which corresponds to a 50% exceedance probability in the fragility curve. A sensitivity simulation was performed to identify disqualified samples and qualified samples within the fragility curve database with greedy algorithm. An analysis of distribution characteristics in the sample space was performed combined with SHAP (Shapley Additive Explanations) and engineering experience. The differences between the fragility curves fitted by this paper's model and the curves provided by disaster assessment software were discussed with the examples of classical RC frame structures in HAZUS. The results indicate that the engineering parameters of the RC frame prototype structure in available Chinese literature are non-uniform. Fragility curves given in current Chinese literature for RC frame structures are mostly credible. The comparison with HAZUS benchmarks demonstrates that the machine learning models show advantages in parametric dimensions and material coverage with more applicablity to single structure. A visualization platform based on research has been developed for evaluating the credibility of existing fragility curves and automatically generating structure-specific fragility curves for RC frame systems rapidly.