And also the LLM version, and LLMZip
However I would say that testing on alice29, enwiki8, text8 data is kinda cheating. Alice in Wonderland and Wikipedia are very likely part of the training data of the LLM models used there.
So I tried on HN comments from a few days ago, extracted from the text column of the public HN bigquery dataset.
Using RWKV v7 0.1B instead of RWKV v4, I get 0.962 bits per byte on alice29, and 1.156 bits per bytes on the HN comments. Still a lot better than 2.826 bits per bytes of xz level 9.
Have you seen this leaderboard of sorts[1], and this proposal to change hutter prize[2]?
I think it's a really clever idea that you could measure an LLM's prediction abilities and language understanding by some sort of held-out compression metric because file sizes are very concrete. They are already beating shannon's numbers using a human prediction for compression, from what i can see.
This is the premise of https://huggingface.co/spaces/Jellyfish042/UncheatableEval
Which implies this would probably also hold for the larger models, which are sadly not included in the leaderboard.