A research team trains four language models (A, B, C, D) of the same architecture but with progressively larger datasets. They plot the final test loss for each model against its dataset size on a graph where both axes are logarithmic. The points for models A, B, and D form a clear, downward-sloping straight line, indicating performance improves predictably with more data. However, Model C, trained on a dataset larger than B's but smaller than D's, has a test loss that is substantially higher than the trend line suggests. Propose a plausible, data-related reason for Model C's anomalous performance and explain your reasoning.