Multi-core fiber optic sensors (MCF) are a promising technology for high-resolution strain monitoring, but their raw output often suffers from drift, nonlinearities, and cross-core inconsistencies. This work presents a laboratory study in which an experimental MCF sensor is evaluated on a steel beam while conventional strain gauges (SG) serve as reference measurements. The objective is to use supervised machine-learning algorithms to detect systematic error patterns in the MCF response and generate a corrected strain signal aligned with the SG reference. During a controlled loading–unloading test, the MCF exhibited a repeatable drift during the recovery phase. Several supervised algorithms were trained to learn the mapping between the MCF raw signal and the SG strain. The learned models successfully reproduced and corrected the device-specific error patterns, producing a calibrated MCF signal with significantly improved accuracy. The results demonstrate that supervised learning provides a powerful data-driven calibration framework for emerging optical fiber sensors. By learning the relationship between the experimental sensor and a well-established reference, supervised models can compensate systematic distortions without relying on analytical assumptions. This contributes to faster deployment and increased reliability of multi-core fiber optics in structural monitoring applications.