JamoFlow reproducibility snapshot v0.1.0
This release accompanies the named preprint:
Causal Whitespace Patching for Korean Byte-Latent Language Models: Quality-Preserving Latency and the Limits of Scale Amplification
Gyeongchan Hwang (Priming Water) · ORCID: https://orcid.org/0009-0007-5840-3274
Archived preprint and reproducibility deposit: https://doi.org/10.5281/zenodo.21973009
All-version concept DOI: https://doi.org/10.5281/zenodo.21973008
Hugging Face research-artifact mirror: https://huggingface.co/datasets/boaz-hwang/JamoFlow-reproducibility
The controlled and free-running measurements show a reproducible latency reduction of approximately 2.5–2.9% at matched quality on the evaluated Apple MPS setup. The prespecified 10% target was not met, and scaling to the evaluated 188.6M-parameter model did not amplify the percentage gain. The release therefore presents a bounded positive systems result and a negative scale-amplification result; it does not claim a general 10% speedup, production readiness, or hardware-general performance.
The arXiv cs.CL submission is awaiting category endorsement. Its identifier will be added without changing the scientific claims or archived assets.
Assets
JamoFlow-paper.pdf— named 9-page preprintjamoflow-arxiv-source.tar.gz— self-contained arXiv source bundlejamoflow-reproducibility-v1.tar.gz— curated code, manifests, evidence summaries, tests, and paper sources
SHA-256
JamoFlow-paper.pdf:393e25d4d204e3e805cba3505267a68508eaf55b2b40ba24795a637cb28f8b3cjamoflow-arxiv-source.tar.gz:ee94268cfc6f28853a7aca33249cede5b20622e565c2e875113d20a72d1eb5bajamoflow-reproducibility-v1.tar.gz:0f48cb350593843958754fa1cddf37b5dd15f642a8e5afdc7f7300681b5bb93b
The release excludes raw corpora, model checkpoints, raw generated outputs, and per-sequence loss arrays. Code is Apache-2.0; the paper is released under CC BY 4.0.