Well, early on there are many weirdly titled pages, and I could imagine that the first 100MB includes all the ’1958 in British Tennis’-style year articles. But intuitively that doesn’t feel like enough to cause bad results.
Nor have any of the articles or theses I’ve read on vandalism detection noted any unusual distributions of vandalism; further, obvious vandalism like gibberish/high-entropy-strings are the very least long-lived forms of vandalism—long-lived vandalism looks plausible & correct, and indistinguishable from normal English even to native speakers (much less a compression algorithm).
A window really does sound like the best explanation, until someone tries out 100MB chunks from other areas of Wikipedia and finds they compress comparably to 1GB.
bzip’s window is 900k, yet it compresses 100MB to 29% but 1GB to 25%. Increasing the memory on 7zip’s PPM makes a larger difference on 1GB than 100MB, so maybe it’s the window that’s relevant there, but it doesn’t seem very plausible to me. (18.5% → 17.8% vs 21.3% → 21.1%)
Sporting lists might compress badly, especially if they contain times, but this one seems to compress well.
Well, early on there are many weirdly titled pages, and I could imagine that the first 100MB includes all the ’1958 in British Tennis’-style year articles. But intuitively that doesn’t feel like enough to cause bad results.
Nor have any of the articles or theses I’ve read on vandalism detection noted any unusual distributions of vandalism; further, obvious vandalism like gibberish/high-entropy-strings are the very least long-lived forms of vandalism—long-lived vandalism looks plausible & correct, and indistinguishable from normal English even to native speakers (much less a compression algorithm).
A window really does sound like the best explanation, until someone tries out 100MB chunks from other areas of Wikipedia and finds they compress comparably to 1GB.
bzip’s window is 900k, yet it compresses 100MB to 29% but 1GB to 25%. Increasing the memory on 7zip’s PPM makes a larger difference on 1GB than 100MB, so maybe it’s the window that’s relevant there, but it doesn’t seem very plausible to me. (18.5% → 17.8% vs 21.3% → 21.1%)
Sporting lists might compress badly, especially if they contain times, but this one seems to compress well.
That’s very odd. If you ever find out what is going on here, I’d appreciate knowing.