CoolFace
Datasetpublic

Potpov/grimoire-figr8

Grimoire — FIGR-8 (processed / final training data) The final, post-processed FIGR-8 training data for the Grimoire SVG pipeline. Shipped processed because regenerating it from the raw FIGR-8 corpus (convert + deepsvg normalize + tokenize) is painful. ⚠️ Split differs from the paper. The paper describes a 427K-sample, 1,000-class, 90/5/5 subset (that split file is datasets/make_font/figr8_final_paper_split.csv in the code repo), but it is not reproducible from code (the class… See the full description on the dataset page: https://huggingface.co/datasets/Potpov/grimoire-figr8.

sourceHugging Facemitupdated 3mo agoView on Hugging Face
0likes39downloads
Dataset Card

Grimoire — FIGR-8 (processed / final training data)

The final, post-processed FIGR-8 training data for the Grimoire SVG pipeline. Shipped processed because regenerating it from the raw FIGR-8 corpus (convert + deepsvg normalize + tokenize) is painful.

⚠️ Split differs from the paper. The paper describes a 427K-sample, 1,000-class, 90/5/5 subset (that split file is datasets/make_font/figr8_final_paper_split.csv in the code repo), but it is not reproducible from code (the class selection / spelling fixes / seed are not in the repo). The split shipped here — and used to train the released checkpoints — is 373,388 samples, 16,486 classes, 80/10/10 (convert_figr8.py, per-class train_test_split(test_size=0.1, random_state=42)).

Contents

  • —svgs_simplified.tar.gz — the deepsvg-normalized SVG corpus (374,992 *.svg), the input for stage-1 (VSQ) training.
  • —figr8_vsq_split.csv — ready-to-use VSQ split (columns simplified_svg_file_path,class,description,split); paths are relative to the extracted svgs_simplified/. 99.6% of SVGs carry their real class + description.
  • —tokenized/tokenized.npy — precomputed VSQ token sequences for stage-2 (ART) training.
  • —tokenized/split.csv — train/val/test split + text; rows index into tokenized.npy via index_in_numpy_array.

Grimoire dataset family — recap

datasetwhat's pushedplug-and-play for VSQ (stage 1)?size
FIGR-8 (this repo)processed SVGs + VSQ csv + ART tokens✅ yes — tar xzf svgs_simplified.tar.gz, then train on figr8_vsq_split.csvsvgs 345 MB · csv 35 MB · tokens 530 MB
MNISTraw PNGs + ART tokens⚠️ needs 1 easy step (PNG→tensors, create_precomputed_mnist.py)png 15 MB · tokens 3.7 GB
Emojiprocessed SAM masks✅ yes — extract, CartoonDataset reads them directly5.7 GB
Cartoons— not pushed (no experiments)——

Usage

  • —Stage 1 (VSQ) — plug and play: tar xzf svgs_simplified.tar.gz and train on figr8_vsq_split.csv (dataset key figr8, mlp decoder, num_segments=4, num_codes_per_shape=2, lseg=5.0, alpha=0.1).
  • —Stage 2 (ART) — ready to train on tokenized/{split.csv,tokenized.npy} with the stage-1 VSQ checkpoint from Potpov/grimoire-checkpoints figr8/vsq.

Re-tokenize with scripts/tokenize_svg_dataset.py if you retrain the VSQ.