To enhance the applicability of standard fiber sensors in building fire scenarios, this study conducted the temperature rise experiments of common single-mode and multi-mode fiber sensors based on the temperature-time curve of typical fire experiments. Moreover, the statistical method was employed to analyze its temperature sensing performance and obtain the fiber sensor suitable for building fire warning. After that, the temperature error correction models were established by nonlinear fitting and machine learning methods. The results show that the temperature measurement stability of the multi-mode fiber sensor is better than the single-mode one. The model built by polynomial fitting performs best above 600 °C, reducing the error to a minimum of 1.5%; and the model built by ANN performs best below 600 °C, reducing the error to a minimum of 3.5%. Furthermore, the generalization of the model was verified by subsequent random temperature offset warming experiments.