SF grids and maps
Collection
Production SF/HIGGS codebooks (27-opt menu) and KV/weight sensitivity maps. • 2 items • Updated
SF grids
The 27 Gaussian VQ codebook grids of the SF codec catalog, corrected 2026-08-29 (fp16 CLVQ sweep + header audit). Each file is self-contained: raw centroids and the measured raw fp16 Gaussian distortion in one package.
Naming: ib<index_bits>_p<p>.sf.
Binary format: [uint32 n][uint32 p][float64 distortion] followed by n*p
float32 centroid values. n = 2^index_bits. Runtime shrinkage correction is
kappa = 1 - distortion / p.
Six high-rate extension rows were added 2026-09-05: ib11_p2 through
ib16_p2 (5.5-8.0 bpw, 2048-65,536 centroids), Gaussian CLVQ-trained in
the 2026-09-04 run. They are staged artifacts pending runtime catalog
integration (catalog menus, schema-3 map regeneration, eval-binary
refresh); no pinned revision binds them yet. They supersede and retire
the former sphere-staging branch.
Model-specific sensitivity maps are published separately as
dlab-cmu/sf-map-<model-name> and collected in SF grids and maps.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support