Boa Constrictor: A Mamba-based Lossless Compressor for Scientific Data

BOA Constrictor is an open-source implementation of a Bytewise Online Autoregressive (BOA) compressor built on Mamba state-space models. It targets lossless compression of High Energy Physics (HEP) datasets. The compressor couples a compact Mamba model (order of a few megabytes) with a parallel range coder to predict and encode bytes in a fully online fashion. On representative ATLAS and CMS datasets, BOA achieves substantially higher compression ratios than a strong classical baseline (LZMA-9), while remaining strictly lossless and respecting the original serialisation. The repository includes: Training and evaluation pipelines for Mamba-based bytewise models A streaming encoder/decoder interface suitable for large HEP files Scripts to reproduce key metrics from the paper (compression ratio, throughput, reliability diagrams, Top-k accuracy, confusion matrices) Example configs for adapting BOA to new datasets CPPJoules wrapper for naive energy estimation Reference C++ implementation for solving the portability issue This Zenodo record provides an archived snapshot of the BOA Constrictor code corresponding to the results in the paper “Boa Constrictor: ML-Enhanced Lossless Compression Algorithms for scientific data”. If you use this software in academic work, please cite both this software record and the associated paper.