ManPla/gauge-orbit-quantization-results

Gauge equivalence does not survive quantisation: sweeps and results

Raw sweeps, end-to-end vectors and reading tables behind Gauge equivalence does not survive quantisation: value-output orbit, norm-product bound, and consequences for rotational methods (Manuel Muñoz Plá, 2026, v2.1).

The layout mirrors the code repository, so this dataset can be dropped over a clone of it and every reading script runs unchanged: https://github.com/mmunozpl/gauge-orbit-quantization (tag v2.1). The code and the two PDFs are also deposited at Zenodo, version DOI 10.5281/zenodo.23068428 under the concept DOI 10.5281/zenodo.22904207.

Second version: eight language models

8,692,320 measurements in weights over eight language models of 0.4 to 8.2 billion parameters, and the end-to-end vectors of every condition.

Path	Rows	What it backs
artifacts/logs/quant_orbita/quant_orbita_v2.csv	1,406,496	regime sweep: identity, per-head and shared orthogonal, general GL
artifacts/logs/quant_kappa/quant_kappa_v2.csv	5,311,488	controlled-conditioning sweep
artifacts/logs/quant_kappa/quant_familia_v2.csv	847,392	original family with norms
artifacts/logs/quant_kappa/quant_diagonal_v2.csv	1,118,208	family of diagonal gauges (F3)
artifacts/logs/quant_kappa/punto_balanceado_v2.csv	8,736	the balanced point, eight models and both bit widths
artifacts/logs/quant_orbita/diagnostico_activaciones_{model}.csv	per cell and gauge	D1, circuit error weighted by activation
artifacts/logs/quant_orbita/diagnostico_salida_{model}.csv	per layer and configuration	D2, error of the attention-block output
artifacts/logs/quant_orbita/e2e_v2_{model}.pt	7 conditions	per-window NLL and per-token argmax of the end-to-end validation on WikiText-2
artifacts/logs/quant_orbita/e2e_extra_v2_{model}_{round}.pt	per round	the same vectors for rounds E2 to E8: Haar draws, layers, sharing, second quantiser, layer by layer and second corpus
artifacts/tables/*.csv		every reading table the paper cites, both versions
artifacts/tables/config_claves_v2.csv	167	configuration keys, repository and revision of the eight models
paper/*.pdf		the two PDFs frozen by tag v2.1, byte-identical to GitHub and Zenodo
sello.json		sha256 manifest of the pieces sealed at v2.1

e2e_v2_qwen3-8b_bf16.pt is the discarded half-precision measurement the paper records as a note on precision.

The eight models

Weights are not redistributed here. They are read from their public repositories at these revisions; Meta's models require accepting their licence.

Model in the paper	Repository	Revision
Pythia-410M	EleutherAI/pythia-410m	9879c9b5f8bea9051dcb0e68dff21493d67e9d4f
Pythia-6.9B	EleutherAI/pythia-6.9b	c0e3eee36dc47af0c49f361c74cfe459c09f7f23
RedPajama-7B	togethercomputer/RedPajama-INCITE-7B-Base	78f7e482443971f4873ba3239f0ac810a367833b
Llama-2-7B	meta-llama/Llama-2-7b-hf	01c7f73d771dfac7d292323805ebc428287df4f9
Llama-3.1-8B	meta-llama/Llama-3.1-8B	d04e592bb4f6aa9cfee91e2e20afa771667e1d4b
Mistral-7B	mistralai/Mistral-7B-v0.3	caa1feb0e54d415e2df31207e5f4e273e33509b1
Qwen2.5-7B	Qwen/Qwen2.5-7B	d149729398750b98c0af14eb82c78cfe92750796
Qwen3-8B-Base	Qwen/Qwen3-8B-Base	49e3418fbbbca6ecbdf9608b4d22e5a407081db4

The paper's results tables and figures abbreviate Qwen3-8B-Base, the pretrained-only checkpoint, as Qwen3-8B, and the files here use the key qwen3-8b.

First version: two models with d_h = 64

Path	Rows	What it backs
artifacts/logs/quant_orbita/quant_orbita.csv	170,016	regime sweep: identity, per-head and shared orthogonal, general GL
artifacts/logs/quant_kappa/quant_kappa.csv	642,048	controlled-conditioning sweep
artifacts/logs/quant_kappa/quant_familia.csv	102,432	re-sweep of the original family with norms
artifacts/logs/quant_kappa/punto_balanceado.csv	1,056	the balanced point, both models and bit widths
artifacts/logs/quant_orbita/aciertos_e2e.pt	7 × 5,000	per-image correctness of every end-to-end condition
artifacts/logs/quant_orbita/predicciones_e2e.pt	7 × 5,000	per-image predicted class (argmax) of every condition

Every sampled gauge is reconstructed exactly by replaying the generator sequence the sweep consumed; the matrices are not stored. The ViT-B/16 checkpoint is the one published with the previous work (https://doi.org/10.57967/hf/9742) and is not duplicated here.

Citation

@software{munozpla2026gaugeequivalence,
  author  = {Muñoz Plá, Manuel},
  title   = {Gauge equivalence does not survive quantisation:
             value-output orbit, norm-product bound, and consequences
             for rotational methods},
  year    = {2026},
  version = {v2.1},
  doi     = {10.5281/zenodo.22904207},
  url     = {https://github.com/mmunozpl/gauge-orbit-quantization}
}