Characterizing the arrhythmogenic substrate is essential for personalized atrial fibrillation management, as structural remodeling induces heterogeneous conduction velocity (CV) slowing and altered anisotropy, promoting wave-break and re-entry. Therefore, accurate estimation of these direction-dependent tissue properties is desirable for risk stratification and planning of targeted interventions. Physics-Informed Neural Network (PINN) architectures, such as FiberNet, offer a data-efficient approach to infer these properties from sparse local activation time (LAT) maps by incorporating the anisotropic eikonal equation as a biophysical constraint. However, previous work showed that the model's performance is challenged by gradually changing fiber orientations and the presence of noise in the input LAT maps. In this work, we present a synthetic 3D benchmark framework to evaluate CV estimation methods under realistic anatomical conditions. The setup utilizes a biatrial manifold with spatially varying, rule-based fiber orientations and region-specific tissue properties, accounting for physiological CVs along and across the preferential orientation of the cardiomyocytes. Using these properties as ground truth, synthetic LAT maps are generated via eikonal forward simulations from multiple pacing sites. To replicate clinical constraints, the LAT maps are augmented with Gaussian noise and localized outliers representing mapping artifacts. With this dataset, we evaluate the robustness of FiberNet for clinical challenges, specifically measurement noise and sparse sampling. To overcome limitations observed in earlier studies, we explore alternative regularization schemes, such as Laplacian smoothing, to enhance estimation fidelity for gradually varying fiber fields. Prediction accuracy is quantified using angular errors for fiber orientation and absolute differences for CV. This framework provides a standardized environment for the objective evaluation of inverse methods, inferring atrial tissue properties from local activation times, establishing a necessary foundation for the translation of such models toward patient-specific clinical applications.