The assessment of corneal fluorescein staining is essential, yet current AI models for Corneal Staining Score (CSS) assessments inadequately identify punctate lesions due to annotation challenges and noise, risk misrepresenting treatment responses through "plateau" effects, and highlight the necessity for real-world evaluations to enhance disease severity assessments. To address these limitations, we developed the Fine-grained Knowledge Distillation Corneal Staining Score (FKD-CSS) model. FKD-CSS integrates fine-grained features into CSS grading, providing continuous and nuanced scores with interpretability. Trained on corneal staining images collected from dry eye (DE) patients across 14 hospitals, FKD-CSS achieved robust accuracy, with a Pearson's r of 0.898 and an AUC of 0.881 in internal validation, matching senior ophthalmologists' performance. External tests on 2376 images from 23 hospitals across China further validated its efficacy (r: 0.844-0.899, AUC: 0.804-0.883). Additionally, FKD-CSS demonstrated generalizability in multi-ocular-surface-disease testing, underscoring its potential in handling different staining patterns.