<p>This work introduces a fully annotated synthetic dataset designed to support machine learning-based estimation of fiber orientation in short-fiber reinforced polymers. The dataset comprises 55,092 high-resolution microscopy-like images generated through a physics-informed simulation pipeline that reproduces the morphological characteristics of real microstructures. Each micrograph contains a non-overlapping arrangement of ellipsoidal fiber cross-sections with controlled distributions of aspect ratio, in-plane orientation and areal fraction. These parameters are calibrated using descriptors extracted from X-ray microscopy of ABS-GF20 specimens processed via large-format material extrusion. For every synthetic image, the corresponding second-order orientation tensor is computed analytically from the exact fiber geometry, providing noise-free ground-truth annotations suitable for supervised learning and rigorous benchmarking. The dataset enables reproducible evaluation of deep learning models for orientation prediction and supports the development of automated workflows for composite characterization and intelligent manufacturing. Comprehensive documentation and the accompanying open-source Python code ensure transparency, facilitate reuse and allow users to extend or regenerate the dataset for materials science, computer vision and data-driven manufacturing research.</p>