Identifying students’ learning styles is an important step in supporting adaptive and personalized learning environments, particularly in higher education contexts. This study investigates the use of a transformer-based language model, DistilBERT, for automated learning style identification based on the Felder–Silverman Learning Style Model (FSLSM). Binary responses from the 44-item Index of Learning Styles (ILS) questionnaire were systematically transformed into descriptive textual representations, preserving the underlying FSLSM decision logic while enabling semantic modeling. Separate DistilBERT classifiers were fine-tuned for the four FSLSM dimensions—Active/Reflective, Sensing/Intuitive, Visual/Verbal, and Sequential/Global. Model training was conducted using three epochs within a stratified five-fold cross-validation framework, with the primary objective of assessing optimization stability and representational consistency rather than predictive superiority. Across all dimensions, training loss decreased monotonically, with the most substantial reductions occurring between the first and second epochs, indicating rapid adaptation of pre-trained representations to the structured learning style descriptions. Differences in convergence behavior across dimensions were observed, reflecting variation in class distributions and response patterns. Near-perfect classification metrics were obtained across multiple folds; however, these results are interpreted as evidence of deterministic consistency between the textual inputs and FSLSM-derived labels rather than independent generalization performance. To examine alignment with human judgment, model predictions were further compared with expert assessments conducted independently using standard FSLSM interpretation guidelines. The resulting Cohen’s Kappa coefficient of 1.0 indicates perfect agreement, confirming that the model faithfully reproduces expert-consistent FSLSM categorizations under controlled conditions. Overall, the findings demonstrate that transformer-based models can reliably encode and recover rule-based learning style constructs from descriptive questionnaire data, supporting their use as computational tools for scalable, consistent learning style analysis, while acknowledging the task's deterministic and framework-dependent nature.