gpress

Hutter Prize campaign on cmix-lex. Target: −1% vs the pending record (0.8735 bpc on enwik9). Two levers: speed (byte-identical, buys model capacity inside the 100 h budget) and capacity (bpc gains that must grow with scale to survive).

now running on h2

connecting…

bytes saved vs trunk — the experiment scoreboard

Each line = one variant. Above zero = smaller output than trunk at that slice. Rising to the right = the gain grows with data — the only shape that survives at enwik9.

bpc vs dataset size — absolute

latest run — per slice

slice	bytes	bpc	ratio	compressed	encode	roundtrip

reference field — enwik8 1m slice

cmix-lex compression pipeline