playra/numeric-conformance-packs

Numeric Conformance Packs (Trinity S3AI / t27)

Version: v2.0 (2026-06-10). License: MIT. Maintainer: Trinity S3AI -- admin@t27.ai -- ORCID 0009-0008-4294-6159. Mirror of: gHashTag/t27 conformance/ and specs/numeric/ subtrees. Companion dataset: playra/numeric-format-catalog (the 80-format ruler this pack instantiates). Linked papers:

• Anchor preprint (GoldenFloat identity): arXiv:2606.05017
• Catalog preprint (84-format ruler): arXiv:2606.09686 -- HF paper page

TL;DR

Bit-exact conformance vectors for the GoldenFloat (GF) family of phi-structured floating-point formats, plus two reference packs (phi_identity, phi_ratio) that exercise the canonical phi^2 + phi^-2 = 3 identity. The packs let any third-party kernel implementer check binary-level conformance with abs_error instead of "looks close in plots".

Quick start

from datasets import load_dataset

# Pull a single pack
gf16 = load_dataset("playra/numeric-conformance-packs", "gf16", split="vectors")
print(gf16.column_names)

# Verify round-trip on one vector
sample = gf16[0]
hex_bits   = sample["bits_hex"]
expected   = sample["f32_value"]
# Your decoder under test:
recovered  = your_gf16_decoder(int(hex_bits, 16))
abs_error  = abs(recovered - expected)
assert abs_error <= sample["tolerance"], f"GF16 conformance violated: {abs_error}"

The full call-and-response contract is documented per pack in manifest.json.

Pack inventory

Config	File	Lines	SHA-256 (8)	Purpose
gf4	gf4.jsonl	1	dd264387	GF4 reference vectors
gf8	gf8.jsonl	1	44e27c75	GF8 reference vectors
gf12	gf12.jsonl	1	682e824d	GF12 reference vectors
gf16	gf16.jsonl	1	a7935733	GF16 reference vectors (frozen silicon)
gf20	gf20.jsonl	34	8bec16f1	GF20 reference vectors
gf24	gf24.jsonl	34	546019d0	GF24 reference vectors
gf32	gf32.jsonl	37	0dbb3a9c	GF32 reference vectors
goldenfloat_family	goldenfloat_family.jsonl	25	a6660f43	Family-level invariants and cross-checks
phi_identity	phi_identity.jsonl	3	88563e4e	phi^2 + phi^-2 = 3 reference points
phi_ratio	phi_ratio.jsonl	1	90a183f4	Canonical phi as f64 with hex tolerance

Full SHA-256 manifest: SHA256SUMS.txt. Machine-readable schema and per-pack manifest: manifest.json.

Provenance

This dataset is a mirror, not a primary source. The primary source of truth is the t27 repository:

• Specs: gHashTag/t27 -> specs/numeric/ (.t27 SSOT files)
• Conformance generators: gHashTag/t27 -> conformance/ + tools/gen_*.py
• Frozen silicon anchor: GF16 is validated end-to-end against the TTGF26a / GF180MCU tape-out (35/35 vectors at 323 MHz, see arXiv:2606.09686 Section 6).

Every JSONL row carries:

• format_name, format_bits, format_spec -- canonical identity
• bit_layout / format_spec -- exact field widths
• seal -- SHA-256 of the row payload for tamper-evidence
• formal_proof (where applicable) -- Coq lemma name in coq/Kernel/*

How this fits the wider ruler

These packs sit at Tier 1 of a three-tier numeric-format corpus:

• Tier 0 SSOT: specs/numeric/formats_catalog.t27 (machine-checked, 83 rows live)
• Tier 1 Portable JSON packs (this dataset)
• Tier 2 Silicon-validated ROM: tt-trinity-corona 80-format ROM + 19 RTL decoders + 5-layer dossier (TTGF26a / GF180MCU tape-out)

The companion dataset playra/numeric-format-catalog carries the full 80-format catalog with cluster, bit-width, and phi-distance metadata.

Considerations for using the data

Recommended uses

• Verifying a third-party GF or phi-arithmetic kernel against a fixed reference
• Cross-checking custom GoldenFloat decoders during silicon bring-up
• Teaching: bit-exact rounding behaviour for non-IEEE phi-structured formats
• Audit and reproducibility studies of phi-structured arithmetic

Out of scope

• IEEE P3109 OFP8 -- cross-walk is documented in p3109_crosswalk.json (mapping only) but not a primary distribution here
• NF4 / NF8 (NormalFloat) -- adjacent quantisation formats, not covered
• MXFP4 / MXFP6 -- OCP mixed-precision wrappers, see arXiv:2606.09686 Section 4 for the cross-walk
• Posit / Takum / LNS -- alternative number systems are catalogued in playra/numeric-format-catalog but no conformance pack here
• Mixed-precision GEMM -- this is a per-format conformance dataset, not an end-to-end kernel benchmark

Known limitations

• Pack size is uneven by design. Several packs ship 1 row (anchor); others ship 34-37 rows (sweep). The number of rows reflects how the format is tested in the kernel, not its importance.
• abs_error only. PCC, FID, and other distributional metrics are explicitly out of scope; this is a pointwise conformance dataset.
• GF dominance. Reflects the maintainer's research focus. For SOTA AI quantisation formats (NF4, MXFP4, posit), this dataset is not a substitute for OCP or open-compute reference implementations.
• Single-language schema. All vectors and manifest are English-only. No localisation.
• No social-bias dimension. Numeric format conformance is not a fairness-sensitive task; if you're using these vectors as part of a downstream model evaluation, fairness considerations live at the model level, not here.

Known biases

• Equal weight per cluster. The pack distribution mirrors the maintainer's research priorities; it is not a representative sample of the AI accelerator format population.
• Phi-structured emphasis. All formats with the goldenfloat tag are designed around the closed-form rule e = round((N-1)/phi^2). This is by construction, not a discovered property.

Citation

@misc{vasilev2026numericrulercatalog,
  title  = {Eighty Formats and a Ruler: A Comprehensive Numeric Format Catalog
            for Phi-Structured and IEEE Floating-Point Arithmetic},
  author = {Vasilev, Daniil and {Trinity S3AI}},
  year   = {2026},
  eprint = {2606.09686},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AR},
  url    = {https://arxiv.org/abs/2606.09686}
}

@misc{vasilev2026goldenfloatidentity,
  title  = {GoldenFloat: A Phi-Structured Floating-Point Family with Closed-Form
            Field Widths from the Identity Phi^2 + Phi^-2 = 3},
  author = {Vasilev, Daniil and {Trinity S3AI}},
  year   = {2026},
  eprint = {2606.05017},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AR},
  url    = {https://arxiv.org/abs/2606.05017}
}

@misc{trinitys3ai2026numericpacks,
  title  = {Numeric Conformance Packs},
  author = {{Trinity S3AI}},
  year   = {2026},
  publisher = {Hugging Face},
  url    = {https://huggingface.co/datasets/playra/numeric-conformance-packs}
}

Maintenance and changelog

Date	Version	Change
2026-04-07	v1.0	Initial 10-pack release
2026-06-10	v2.0	Updated with audit-disclosed schema, SHA-256 manifest, anchor preprints
2026-06-13	v2.1	Added catalog preprint cross-link (arXiv:2606.09686), README upgrade per FAIR / Datasheets / Data Cards

Update cadence. No fixed cadence. New packs are added when (a) a new GF rung is verified, or (b) a new conformance contract is requested via gHashTag/t27 issues. Existing rows are immutable; if a format spec changes, the row is replaced and the SHA-256 manifest tracks the diff.

Contact. admin@t27.ai -- async-only, ASCII-only.

Machine-readable metadata

• Croissant JSON-LD: croissant.json at repo root (MLCommons spec 1.0)
• YAML metadata: at the top of this file, machine-parseable per HF Hub spec
• SHA-256 manifest: SHA256SUMS.txt
• P3109 cross-walk (read-only mapping): p3109_crosswalk.json

Identity stack

• Maintainer: Vasilev Daniil (Trinity S3AI)
• Email: admin@t27.ai (sole canonical contact)
• ORCID: 0009-0008-4294-6159
• GitHub: gHashTag
• HF org: trinity-s3ai
• Anchor identity: phi^2 + 1/phi^2 = 3