Taco

Taco is a proof-of-concept lossless data compressor that leverages a GPT-2 language model to predict data and stores only the unpredictable parts via range coding. It uniquely guarantees that compression and decompression produce the exact same output bytes across different hardware architectures and platforms by using integer-only arithmetic. While slower than conventional compressors, Taco focuses on reproducible cross-machine compression results and serves as an experimental tool for research in language-model-based compression.