In this Letter, a knowledge-distillation-inspired cascaded multi-modulus algorithm-based deep neural network (KD-CMMA-DNN) scheme is proposed to achieve a high-performance semi-supervised equalizer in intensity-modulation and direct detection (IMDD) systems. In this scheme, a pretrained teacher model is utilized to assist the CMMA model by a specially designed distillation loss function, enabling the model to exhibit superior performance compared to a typical blind CMMA equalizer. The proposed KD-CMMA-DNN equalizer demonstrates significant effectiveness in the O-band PAM-4 IMDD system. We experimentally verified that the use of a KD-CMMA-DNN equalizer enabled the O-band 50-Gb/s PAM-4 transmission over a 25-km standard single-mode fiber to reach the 7% hard-decision forward error correction threshold. Meanwhile, the proposed scheme can eliminate the need for labeled data, significantly reducing system costs without performance degradation in comparison with the supervised DNN equalizer.