clar-enwik

Six fixed 20 MB byte slices of enwik8 and enwik9, used by continual-learning-autoresearch to score online learners: an algorithm reads the stream one byte at a time, predicts the next byte, and pays its log-loss (cumulative prequential loss, reported in bits per byte).

The files are raw bytes, exactly as they appear in the source: no tokenization, no decoding, no cleaning. Each slice is identified by its SHA-256; the benchmark's harness checks it on every load.

Files

file	source	offset (bytes)	length	sha256	use
dev/enwik8_10M.bin	enwik8	10,000,000	20,000,000	c70e11eb…4318ad	dev
dev/enwik8_40M.bin	enwik8	40,000,000	20,000,000	5d858d96…dfca30	dev
dev/enwik8_70M.bin	enwik8	70,000,000	20,000,000	647cae92…9125d7	dev
report/enwik9_200M.bin	enwik9	200,000,000	20,000,000	abde340a…9591ba	report only
report/enwik9_500M.bin	enwik9	500,000,000	20,000,000	44e218e5…f20ab6	report only
report/enwik9_800M.bin	enwik9	800,000,000	20,000,000	fad93d33…cec36e	report only

Full hashes are in SHA256SUMS (sha256sum -c SHA256SUMS).

Each slice is source[offset : offset + length], where source is the uncompressed enwik8 or enwik9 file from https://mattmahoney.net/dc/enwik8.zip / enwik9.zip. enwik8 is the first 10^8 bytes of enwik9, so the report/ slices (offsets ≥ 200,000,000) don't overlap the dev/ slices. A run uses the first N bytes of a slice.

dev vs report only. Tune on dev/. The report/ slices are public on purpose: they are for reporting how a method transfers to unseen text, and tuning on them defeats that. Nothing enforces it; papers that tune on report slices can be exposed by others rerunning their experiments on new data.

The raw source: enwik9

enwik9 is the uncompressed source file itself (1,000,000,000 bytes, sha256 159b85351e5f76e60cbe32e04c677847a9ecba3adc79addab6f4c6c7aa3744bc), unchanged from https://mattmahoney.net/dc/enwik9.zip. enwik8 is its first 10^8 bytes. Every slice above is enwik9[offset : offset + 20,000,000]; the benchmark now defines its streams as byte ranges of this file, so the slice files are kept only for older references.

Source and license

enwik8 and enwik9 are Matt Mahoney's excerpts of the English Wikipedia XML dump of 3 March 2006, used by the Large Text Compression Benchmark and the Hutter Prize. The text is Wikipedia content, available under the Creative Commons Attribution-ShareAlike 3.0 license (and the GFDL), with attribution to Wikipedia's contributors. These slices are redistributed unchanged under the same terms.

Reproduce the slices

import hashlib, io, urllib.request, zipfile

def cut(source, offset, length=20_000_000):
    z = zipfile.ZipFile(io.BytesIO(urllib.request.urlopen(
        f"https://mattmahoney.net/dc/{source}.zip").read()))
    with z.open(source) as f:
        f.read(offset)
        return f.read(length)

data = cut("enwik8", 10_000_000)
print(hashlib.sha256(data).hexdigest())  # c70e11ebd1e320a81986f7f960d3a120c90bca6bd1d4bc45f8a14a35624318ad