Loading...
Loading...
Contributors:
Dataset URI:
10.5281/zenodo.23323548Machine-readable measurements, provenance documentation and figures behind the manuscript The Discipline of Sparse Attention: A Budget-Matched Validation from 350M to 27B.What version 2 adds. Version 1 deposited 38 JSONL files. Version 2 adds the measurements that support the revised manuscript: the two balanced n=40 RULER batteries (2600 rows each), the external-pattern battery with the platform recorded per row, the paired-needle re-run behind the exact McNemar test, the decode-throughput and block-granularity benchmarks, the 20-seed SmolLM2 confidence-interval panel and the 4B budget curve; together with the four figures, the provenance manifest, the environment manifest and the execution log that evidences the notebook-tier platform. The 38 files of version 1 are reproduced byte-for-byte and their checksums are unchanged.Start with MANIFEST.md. It maps every numbered table and figure of the paper to the data file and the script that produced it, and states for each row how the attribution was established: V (verified by recomputation or by an exact match of the file's categorical vocabulary with the table), A (attributed from the manuscript itself or from the recorded procedure of the campaign) or D (definitional or derived). It also documents the known artefacts of the deposited files and the naming convention used here.File naming. Files are stored at a single level, so a file that lives in a subdirectory of the source repository is deposited with its path separator replaced by a double underscore: results/old_bug/ruler_battery.jsonl is deposited as old_bug__ruler_battery.jsonl. Files at the top level of results/ keep their name unchanged.Code and licensing. The measurements in this record are licensed under CC-BY-SA-4.0. The code that produced them — the attention interceptor and harness, one driver per experiment, the analysis scripts that assemble the published aggregates, and the figure script — is licensed under GPL-3.0 and is maintained in the accompanying software repository listed under related identifiers.Comparability. Four platforms produced these numbers. Wall-clock and peak-VRAM figures are comparable only within a platform, never across platforms; environment_manifest.json records each platform, its exact software stack, what it produced, and the evidence for that attribution.Note on version 1. Version 1 remains available under its own DOI and was licensed CC-Zero. That waiver is irrevocable and continues to apply to those 38 files as deposited there; the CC-BY-SA-4.0 licence stated here governs this version.
Keyword(s):
attention sparsity; efficient inference; hybrid transformers; KV cache; long context.
Grants and Funding:
No Grants or Funding Data Available
Associated Publication(s):
No Associated Publication Data Available
Supplementary Data:
No Supplementary Data Available
Related Dataset(s):
No Related Dataset Data Available
The National Library of Medicine (NLM) provides information about biomedical datasets. Inclusion on an NLM Web page does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: NLM Web Policies (nih.gov)