Training AI models requires ingesting massive amounts of training data. We define Parametric Matching (PM), a grammar-based compression technique that parses structured text into abstract syntax trees, replaces format patterns with compact tokens, and moves numerical parameters into homogeneous streams of tokens that can additionally be bit compressed using LZMA. The results are input streams that are equivalent to the original interleaved text and data in formats such as SVG, 3D-models such as OBJ, PDF, but highly compressed and much more amenable to processing. PM has many advantages. Loading data is much faster (a factor of 20 for 3D models, and 100 to 200 times faster to load). Just as important, by preprocessing the semantics, ambiguous cases can be resolved once, and can be turned into embedding vectors more efficiently, and potentially more accurately. Our experiments with SVG, G-Code, OBJ and PDF files show large compression. The results in Table 1 show that the PM algorithm can create data objects that are not only more compressible but far faster to load because the binary format can be instantly used, in this case to load OpenGL and render without parsing ASCII data. When compressing a binary object that is half the size of the original ASCII representation, by definition LZMA takes half the time to compress because there are fewer bytes to process.

Modern general-purpose compressors can compress a wide variety of files but do not achieve high compression ratios on files that contain short sequences of delimiters with interleaved numeric data and generally with interleaved data where each sequence is not well correlated to the previous bytes. We demonstrate Parametric Matching (PM), which vastly improves the compression of various structured languages, including PDF, SVG, and G-code files. By de-interleaving and coalescing delimiters and storing data as delta-encoded, discretized binary, compressions of a factor of 10 or more are possible. A Python prototype compresses files to a binary representation, which is then compressed using Lempel-Ziv-Markov (LZMA) to efficiently store the binary tokens in a minimal number of bits. Table 1 shows a ratio of 6 for PDF files containing only text, which are first parsed, and recompressed using PM. For SVG, we demonstrate a factor of 8 to 10 for files including a randomized spiral and a US county map. For the G-code, we compressed the Statue of Liberty, demonstrating that even when the layers are different, a high degree of compression can be achieved. Times are all less than 250ms, even in our Python prototype.