An Empirical Comparison of Methods for Quantifying the Similarity of Numeric Datasets

This project is an academic empirical study rather than a software product or infrastructure tool. With 0 stars and 3 forks at 1 day old, it represents a 'snapshot' of research. Its primary value is the 'neutral comparison' of existing methods (e.g., Kolmogorov-Smirnov, Maximum Mean Discrepancy, Wasserstein distance), which is useful for practitioners in synthetic data and ML monitoring but lacks any technical or economic moat. Defensibility is minimal as the methods evaluated are well-documented in statistical literature and existing libraries (SciPy, SDV). Frontier labs have little interest in competing with a benchmark, though they utilize the underlying metrics for model evaluation. The risk of displacement is moderate only because benchmarks age as newer, more robust metrics (like those based on neural embeddings) gain traction over traditional statistical tests. It serves as a valuable reference implementation for selecting the right metric for specific data distributions, but it is not a standalone platform.