A speedrun for training a small language model, with a meter where the clock would be.
Every tensor operation costs ticks, priced from its shape alone. An entry gets 40 billion of them, so a laptop and a server run the same program to the same reading, with scores that agree far below the record margin. Hardware only buys finishing sooner.
The current record's code, unchanged, at a 2 billion tick budget on a 4 MiB excerpt. It downloads under 20 MB, all from this site, and stays in this tab.
The record table
Click a row for its code, every seed, and the written review. The rule for a new record: beat the holder by more than 0.002 bits per byte at p < 0.01, then do it again on 8 new seeds, then get a written review that explains why the gain is real.
Check it yourself
This reruns the current record on 2 fresh seeds on the test split, and prints your number next to the stored one.
git clone https://github.com/abhaymettu/odometer && cd odometer
python3 -m venv .venv && .venv/bin/pip install numpy==2.5.4
.venv/bin/python reproduce.pyWhat a tick buys
A tick is about a nanosecond of one laptop core, priced so that an entry can't get cheap work by picking the op the list underprices. Change the budget and the recipes keep their order at the top. The tuned recipes here are not retuned at other budgets, so a line that bends the wrong way is a recipe outgrowing its settings.