hai-minh-son/processed-fas-data

Processed FAS Data

Preprocessed face anti-spoofing (FAS) tensors serialized as PyTorch archives. The original upload contains no documentation; this card describes the observed repository structure only. Claims about content semantics are marked unverified.

Repository structure

Folder	Contents	.pt files
adv_catibins/	metadata.pt + batch_files/batch_0000.pt … batch_0030.pt	32
catibins/	metadata.pt, metadata_temp.pt + batch_files/batch_0000.pt … batch_0021.pt	24
celeba/	metadata.pt + batch_files/batch_0000.pt … batch_0033.pt	35

Total: 91 .pt files, ≈103 GB.

Content (unverified)

• The folder names suggest: a CaTi FAS data partition split into bins (catibins), an adversarial variant (adv_catibins), and a CelebA-derived split (celeba).
• metadata.pt most likely holds labels/protocol information for the corresponding batch_files/ shards. This has not been verified — inspect metadata.pt with torch.load to recover the schema.

Loading

from huggingface_hub import hf_hub_download
import torch

path = hf_hub_download("hai-minh-son/processed-fas-data", "celeba/metadata.pt",
                       repo_type="dataset")
obj = torch.load(path, map_location="cpu", weights_only=True)

Pickle warning: .pt files are pickle-based archives. Only load them from sources you trust; prefer weights_only=True where compatible.

Access: this repository is gated (manual approval) — submit an access request and wait for approval before downloading.

License

No license has been specified (license: unknown). Underlying face imagery is subject to the licenses of its source datasets (e.g. CelebA is licensed for non-commercial research). Users are responsible for ensuring compliant use.

Intended use & limitations

• Research on face anti-spoofing / presentation-attack detection.
• No train/validation split documentation; no class-label mapping included in the card — recover from metadata.pt.
• catibins/metadata_temp.pt appears to be a leftover temporary file; treat with caution.

Citation

@misc{processed_fas_data,
  title        = {Processed FAS Data},
  author       = {{hai-minh-son}},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/hai-minh-son/processed-fas-data}}
}