Zstandard (zstd), with a properly trained dictionary, could do well against larger (than alice29) texts. I doubt you'd see ~0.4 bpb, but 1-1.5 seems achievable.

Zstandard (zstd), with a properly trained dictionary, could do well against larger (than alice29) texts. I doubt you'd see ~0.4 bpb, but 1-1.5 seems achievable. And it would be ~100% accurate with no GPU or other overhead necessary. A quick experiment using a small selection of truncated texts yielded 1.8 bpb against alice29. I imagine other texts could do better.