Hugging Face's logo

Datasets:
playra
/
numeric-format-catalog

Languages:
English
Size:
n<1K
ArXiv:
Tags:
numeric-formats
floating-point
conformance
audit-disclosed
License:
Dataset card Files
xet
 Community
numeric-format-catalog / README.md
playra's picture
playra
v2: count drift audit -- 81 formats with full disclosure
f50072c verified
preview code
|
raw
history blame contribute delete
3.51 kB
metadata 
license: cc-by-4.0
language:
  - en
tags:
  - numeric-formats
  - floating-point
  - conformance
  - audit-disclosed
size_categories:
  - n<1K

numeric-format-catalog v2 (audit-disclosed)

Version: v2.0 (2026-06-10) -- supersedes v1 (count=77, withdrawn).

Count: 81 formats across 13 clusters.

Honest count-drift disclosure

This dataset is shipped with a known count-drift across four artefacts:

Source Count Status
arXiv:2606.09686 Table 1 (paper #3) 84 aspirational, MXFP6 split + standalone E8M0
specs/numeric/formats_catalog.t27 SSOT raw // CATALOG: lines 83 live ground truth
THIS DATASET (codegen LIVE re-run) 81 parser drops gf512 + gf1024 on bias-formula int() cast
gen/numeric/formats_catalog.json (committed in repo) 77 stale, pre-#1051 snapshot
v1 of this dataset (uploaded 2026-06-10 AM) 77 inherited stale gen

Why 81 and not 83 or 84: the codegen tools/gen_formats_catalog.py cannot evaluate bias formulas of the form 2^N-1, which are the canonical bias values for gf512 and gf1024. Those two rows are silently dropped. We ship what the codegen actually produces, not a cherry-picked number. Patch tracked in t27#1064.

Why ship 81 today instead of waiting for the patch: because v1 shipped 77 yesterday. Replacing a stale 77 with an honest 81 + full audit is closer to the truth than leaving 77 up while the patch lands.

What's in the dataset

  • 81 row entries, each carrying: id, name, bits, s (sign bits), e (exponent bits), m (mantissa bits), bias, phi_distance, storage, cluster, status, standard, use_case, source
  • 13 cluster groupings: IEEE754 binary, IEEE754 decimal, MLLowPrecision, GoldenFloat, Posit/Unum III, OCP MX, LNS, IntegerFixed, HistoricalVendor, Theoretical, Compression, Extended, QuantTuned
  • Source SSOT: gHashTag/t27 specs/numeric/formats_catalog.t27 HEAD 6ecad30
  • Regen output SHA-256: 546ba6fff1daec0b64483f599778ab63a9410405592c2e9dbc9d90dbb63bc48a

Known gaps vs paper #3 (arXiv:2606.09686)

Cluster Paper says This dataset Delta
GoldenFloat 16 16 0 (both pre-#1051; current SSOT has 22, of which gen lifts 20)
OCP MX 5 3 -2 (no MXFP6 E2M3/E3M2 split; no standalone E8M0 block)
MLLowPrecision 8 7 -1 (NF4 placement)
Total 84 81 -3

Reconciliation

  • t27#1064 -- tracking issue, five-step plan: parser patch -> regen -> CI invariant -> HF re-upload v3 -> arXiv erratum if needed
  • Expected v3 ship count after parser patch: 83
  • Expected v4 ship count after MXFP6/E8M0 reconciliation with paper: 84 (or paper erratum to whatever codegen produces)

Citation

If you use this dataset, please cite the paper it accompanies:

@misc{vasilev2026numeric,
  title={An 84-Format Numeric Catalog with Bit-Exact Conformance Vectors},
  author={Vasilev, Dmitrii},
  year={2026},
  eprint={2606.09686},
  archivePrefix={arXiv},
  primaryClass={cs.MS}
}

Note: paper claims 84; this dataset ships 81 due to the parser bug. See "Honest count-drift disclosure" above.

Provenance

  • Uploader: playra (Dmitrii Vasilev, ORCID 0009-0008-4294-6159)
  • Affiliation: Trinity S3AI
  • Repo: gHashTag/t27
  • License: CC BY 4.0
  • Generated: 2026-06-10 +07 from HEAD 6ecad30