[HN]score: 0.11
Overfitted a 900KB Transformer to Compress a 100MB CSV into 7MB
June 23, 2026
A 900KB transformer overfitted to a single file acts as a per-byte predictor, with its probability outputs fed into an arithmetic coder. On a 100MB NYC taxi CSV this achieves 7MB (0.5 bits/byte); on 100MB of enwik9 it reaches 21MB (1.68 bits/byte). The compressed output requires the model weights plus arithmetic-coded bytes, so decompression needs the stored model.