With the rapid advancements in the field of fluorescent dyes, accurate prediction of optical properties and efficient retrieval of dye-related data are essential for effective dye design. However, there is a lack of tools for comprehensive data integration and convenient data retrieval. Moreover, existing prediction models mainly focus on a single property of fluorescent dyes and fail to account for the diverse fluorophores and solutions in a systematic manner. To address this, we proposed Fluor-predictor, a multitask prediction model for fluorophores. This study integrates multiple dye databases and develops an interpretable graph neural network-based multitask regression model to predict four key optical properties of fluorescent dyes. We thoroughly examined the impact of factors such as data quality and the number of solvents on model performance. By leveraging atomic weight contributions, the model not only predicts these properties but also provides insights to guide structural modifications. In addition, we compiled and built a comprehensive database containing 36,756 records of fluorescence properties. To address the limitations of existing models in accurate prediction of Xanthene and Cyanine dyes, we then compiled 1148 Xanthene dye records and 1496 Cyanine dye records from the literature, comparing direct training with transfer learning approaches. The model achieved mean absolute errors (MAE) of 11.70 nm, 15.37 nm, 0.096, and 0.091 for predicting absorption wavelength (λ