This study presents a novel approach to automated netlist generation from impedance spectra using a multi-task transformer model. A netlist is a textual representation of a component’s equivalent circuit topology and the parameter values. Traditionally, designers begin with an impedance measurement and iteratively reverse-engineer a circuit netlist that approximates the desired behavior. This process is often time-consuming and computationally expensive. To facilitate model training and validation, a synthetic dataset was developed that captures realistic behaviors of passive components. The dataset was designed to provide broad coverage across various component value ranges and diverse impedance profiles, enhancing its utility as a benchmarking resource. To tackle the problem, a multi-task transformer model is proposed. This model simultaneously predicts the circuit topology (a classification task) and component values (a regression task). Given the inherent difficulty of multi-task learning in this context, several gradient optimization strategies were explored. Among these, GradVac emerged as the most effective, yielding the lowest validation loss (MSE = 0.0074) and producing impedance spectra closely aligned with ground truth. Both GradVac and CaGrad demonstrated efficacy in mitigating challenges arising from task imbalance, such as differing magnitudes, conflicting gradient directions, and high-curvature optimization paths. The effectiveness of the proposed approach was validated using the newly developed dataset, where the model demonstrated strong performance. These results underscore the potential of transformer-based architectures for rapid and accurate equivalent circuit modeling and netlist generation.