jhtyip/cytransformer-frst-datasets

CYTransformer — FRST training datasets

Favorable 4-dimensional reflexive polytopes paired with Fine, Regular, Star Triangulations (FRSTs) — the combinatorial data behind smooth Calabi–Yau threefolds — pre-split into train / validation / test.

These are the datasets behind "Transforming Calabi–Yau Constructions: Generating New Calabi–Yau Manifolds with Transformers" (arXiv:2507.03732), released for AICY.

Code: github.com/jhtyip/cytransformer · Weights: jhtyip/cytransformer-frst-models

Contents

One folder per configuration. 9+1 means 9 polytope vertices plus the origin; the paper's N_vert counts the origin, so folder 9+1 is the paper's N_vert = 10.

Folder	h1,1	Train	Val	Test	Size	Sampling
9+1	5	3,897 / 45,416	500 / 5,763	500 / 5,535	20 MB	exhaustive
10+1	6	14,608 / 513,973	1,000 / 36,213	1,000 / 34,095	242 MB	exhaustive
11+1	7	42,221 / 5,266,151	3,000 / 347,784	3,000 / 376,398	2.88 GB	exhaustive
12+1	8	7,425 / 3,751,111	1,000 / 503,231	1,000 / 488,398	2.60 GB	exhaustive
13+1	9	16,727 / 4,386,161	1,000 / 260,805	1,000 / 258,880	3.03 GB	fast sampler, ≤300 per polytope
14+1	10	16,486 / 4,815,210	1,000 / 291,089	1,000 / 291,313	3.72 GB	fast sampler, ≤300 per polytope

Train / val / test columns are polytopes / pairs. Every count above was measured from the files, not taken from filenames. Splits are per-polytope, so no polytope appears in more than one split.

Format

Six files per folder — {polys,triangs}_{train,val,test} — as plain JSON lists of [POLYID, STRING] records. Row i of the polys file pairs with row i of the triangs file, so a polytope is repeated once per triangulation.

polys:   [3081, "{{0,1,0,0},{1,0,0,0},{-3,-2,-1,-1},{0,0,0,1},{0,0,1,0},{-2,0,-1,0},{-1,0,1,-1},{-1,1,0,0},{-2,-1,0,-1}}"]
triangs: [3081, "{{0,1,2,5},{0,1,2,8},{0,1,3,4},{0,1,3,5},{0,1,4,6},{0,1,6,8},{0,2,5,8},...}"]

A polytope is its list of vertices in ℤ⁴ (DRESVERTS). A triangulation is its list of 4-simplices, each given as indices into that vertex list. Polytopes are favorable — vertex count equals lattice-point count minus the origin.

Because each polytope is duplicated once per triangulation, these files compress extremely well (the polys files by more than 100×).

Usage

from huggingface_hub import hf_hub_download
p = hf_hub_download("jhtyip/cytransformer-frst-datasets", "9+1/9+1_polys_train.json", repo_type="dataset")
t = hf_hub_download("jhtyip/cytransformer-frst-datasets", "9+1/9+1_triangs_train.json", repo_type="dataset")

They feed cyt-train directly — the files are already split, so no preparation step is needed.

Provenance

Generated with CYTools from the Kreuzer–Skarke database. Configurations 9+1–12+1 enumerate triangulations exhaustively; 13+1 and 14+1 use a fast sampler capped at 300 triangulations per polytope, since exhaustive enumeration is infeasible at that size.

Published filenames drop the polytope counts that appeared in the original working filenames — those counts referred to scan ranges and configured limits rather than file contents, and did not match the data. The measured counts are in the table above.

Citation

@article{Yip:2025hon,
  author  = {Yip, Jacky H. T. and Arnal, Charles and Charton, Fran\c{c}ois and Shiu, Gary},
  title   = {Transforming Calabi-Yau Constructions: Generating New Calabi-Yau Manifolds with Transformers},
  journal = {arXiv e-prints},
  eprint  = {2507.03732},
  year    = {2025},
  doi     = {10.48550/arXiv.2507.03732}
}

MIT licensed. Department of Physics, University of Wisconsin–Madison.