Quantum Neural Networks (QNNs) are a promising new intersection between classical machine learning and quantum computing. Recent advancements have shown that using entangled training data reduces the amount of data necessary to train a QNN with minimal risk. However, it is not yet fully understood how an increase in degree of entanglement influences the trainability of these QNNs. For this reason, our research aims to analyze the effect of increasing degrees of entanglement and sizes of training sets, as well as different linear structures, such as orthogonal or linearly dependent training samples, on the shape and trainability of QNN loss landscapes. In our experiments, we sample loss landscapes obtained from QNN loss functions for various compositions of training samples. We then analyze the shape of the loss landscapes using multiple roughness metrics. Our findings include correlations between the entanglement, size, and linear structure of the training data and the shape of the corresponding loss landscapes. Most notably, for an increase in degree of entanglement, the loss landscapes show a noticeable decrease in roughness. An increase in training data size has similar results for almost all data structures except for linearly dependent samples, which were significantly less affected. While a smoother landscape can hint at a reduction of local minima and saddle points, this roughness decrease likely implies other training obstacles, such as barren plateaus or narrow gorges. The insights gathered by our experiments can help develop new effective and efficient training strategies for QNNs.