spidy__

I built an experiment that uses an overfitted transformer and arithmetic coding to compress individual files.

Instead of training the model to generalize, I train a 900KB transformer to memorize a single file and predict the next byte. Those predictions are fed into an arithmetic coder to produce the compressed output.

On a 100MB NYC taxi CSV, it compresses to about 7MB (~0.5 bits/byte). On a 100MB slice of enwik9, it compresses to about 21MB (~1.68 bits/byte).

It's pretty slow right now (roughly 20–30 minutes of training and 45 minutes each for compression and decompression on my AMD 7800XT).

Checkout the repo - https://github.com/samyak112/pym-particles