Datasets:

playra
/

numeric-conformance-packs

benchmark_metrics dict	best_phi_approximation bool	created_at timestamp[s]	description string	format_bits int64	format_family string	format_name string	format_spec dict	is_primary bool	phi_distance float64	sacred_ok bool	schema_version int64	seal string	test_vectors list	updated_at timestamp[s]	validated_at timestamp[s]	vector_name string	verdict string	version string
{ "memory_savings": 0.625, "mse_vs_fp32": null, "speedup_vs_fp32": null, "use_cases": [ "high_precision_quantization", "attention_matrices", "critical_path_weights" ] }	true	2026-04-06T15:54:08	GoldenFloat12 Conformance Vectors — NUMERIC-STANDARD-001 (Agent 12)	12	GF12	GF12	{ "exp_bias": 7, "exp_bits": 4, "exp_mant_ratio": 0.5714285714, "mant_bits": 7, "memory_ratio_vs_fp32": 0.375, "sign_bits": 1 }	false	0.046605	true	2	sha256:2f3e1dbb4fc3c235f0a08067294d33844976f2787a9dc72a906eaf4dd5acd0df	[ { "description": "Positive zero encodes to 0x000", "expected": { "decoded": 0, "raw": 0, "tolerance_abs": 0 }, "input": { "value": 0 }, "name": "zero_positive", "verdict": "CLEAN" }, { "description": "1.0 encodes to exp=7, mant=0", "expected": { ...	2026-04-06T15:54:08	2026-04-06T15:54:08	gf12 Vectors	CLEAN	1.0

Trinity S3AI Numeric Conformance Packs

Bit-exact JSON vectors for the GoldenFloat (GF4..GF32) family and the phi-identity invariants used by the Trinity S3AI stack. Each pack is a portable, language-agnostic conformance contract: any implementation that consumes a pack and produces matching outputs is bit-exact with the canonical reference.

Source of truth

This dataset is a discoverability mirror. Updates flow GitHub -> Zenodo -> Hugging Face (one-way).

GitHub (primary): gHashTag/t27/conformance
Zenodo DOI: 10.5281/zenodo.19456875
Anchor preprint: arXiv:2606.05017

Configurations

Config name	What it covers
`gf4`	GF4 (4-bit GoldenFloat) reference vectors
`gf8`	GF8 reference vectors
`gf12`	GF12 reference vectors
`gf16`	GF16 (default) reference vectors
`gf20`	GF20 reference vectors
`gf24`	GF24 reference vectors
`gf32`	GF32 reference vectors
`goldenfloat_family`	Cross-family aggregate vectors (GF4..GF32 together)
`phi_identity`	phi^2 + 1/phi^2 = 3 anchor invariants
`phi_ratio`	phi-ratio derived constants

Load

from datasets import load_dataset

ds = load_dataset("playra/numeric-conformance-packs", "gf16", split="vectors")
print(ds[0])

Row schema

Each JSONL file begins with one _row_type: "meta" row containing top-level metadata from the canonical pack (format family, version, claim status, etc.), followed by zero or more _row_type: "vector" rows -- each a single test vector with input bits and expected output bits.

Honest scope (what is NOT here yet)

This first cut covers the GoldenFloat family + phi-identity invariants. Conformance packs for FP8 (E4M3/E5M2), MXFP4, bf16, posit family, and others exist as catalog rows in the companion numeric-format-catalog but their bit-exact vector packs are still being curated for export to HF. Tracking under gHashTag/t27.

Verification

Reproduce SHA-256 over each pack:

sha256sum *.jsonl
# should match SHA256SUMS.txt

The same hashes appear in manifest.json next to source-of-truth links.

Citation

@misc{vasilev2026goldenfloat,
  author = {Vasilev, Dmitrii},
  title  = {GoldenFloat: a minimal numeric anchor at phi^2 + 1/phi^2 = 3},
  year   = {2026},
  eprint = {2606.05017},
  archivePrefix = {arXiv},
}

@dataset{vasilev2026t27,
  author = {Vasilev, Dmitrii},
  title  = {Trinity S3AI / t27 conformance corpus},
  year   = {2026},
  doi    = {10.5281/zenodo.19456875},
}