spaicom-lab/semasia-celeba

Latents for celeba (timm)

   

This repository hosts precomputed latent representations (embeddings) extracted from timm image-classification backbones on celeba, released as part of SEMASIA — a large-scale resource for studying semantic communication, cross-model latent space alignment, and explainability. Each config corresponds to a single model; only that model's Parquet files are read on load_dataset.

Usage

Load with datasets and convert to torch:

from datasets import load_dataset
import torch

ds = load_dataset(
    "spaicom-lab/semasia-celeba",  # repository  →  which benchmark
    "aimv2_1b_patch14_224.apple_pt",  # config      →  which model
    split="test",  # split       →  which partition
).with_format("torch")

embeddings = torch.vstack(list(ds["embedding"]))  # (N, d)
celeb_id = torch.tensor(ds["celeb_id"])  # (N,)

Or read the Parquet files directly with polars:

import polars as pl

df = pl.read_parquet(
    "hf://datasets/spaicom-lab/semasia-celeba/test/aimv2_1b_patch14_224.apple_pt/*.parquet"
)

embeddings = df["embedding"].to_numpy()  # shape (N, d)
celeb_id = df["celeb_id"].to_numpy()  # shape (N,)

Fields

Columns available in each Parquet file for this dataset:

Field	Description
id	Row index within the shard (unique per split/model, not across models); the row order matches the original flwrlabs/celeba split, so id can be used to map a row back to its source sample.
model_name	timm model that produced this row's embedding (constant within a config).
embedding	Precomputed latent representation extracted by the model (dimensionality depends on config; see model registry).
celeb_id	Original dataset field, copied as-is from the source dataset.
5_o_Clock_Shadow	Original dataset field, copied as-is from the source dataset.
Arched_Eyebrows	Original dataset field, copied as-is from the source dataset.
Attractive	Original dataset field, copied as-is from the source dataset.
Bags_Under_Eyes	Original dataset field, copied as-is from the source dataset.
Bald	Original dataset field, copied as-is from the source dataset.
Bangs	Original dataset field, copied as-is from the source dataset.
Big_Lips	Original dataset field, copied as-is from the source dataset.
Big_Nose	Original dataset field, copied as-is from the source dataset.
Black_Hair	Original dataset field, copied as-is from the source dataset.
Blond_Hair	Original dataset field, copied as-is from the source dataset.
Blurry	Original dataset field, copied as-is from the source dataset.
Brown_Hair	Original dataset field, copied as-is from the source dataset.
Bushy_Eyebrows	Original dataset field, copied as-is from the source dataset.
Chubby	Original dataset field, copied as-is from the source dataset.
Double_Chin	Original dataset field, copied as-is from the source dataset.
Eyeglasses	Original dataset field, copied as-is from the source dataset.
Goatee	Original dataset field, copied as-is from the source dataset.
Gray_Hair	Original dataset field, copied as-is from the source dataset.
Heavy_Makeup	Original dataset field, copied as-is from the source dataset.
High_Cheekbones	Original dataset field, copied as-is from the source dataset.
Male	Original dataset field, copied as-is from the source dataset.
Mouth_Slightly_Open	Original dataset field, copied as-is from the source dataset.
Mustache	Original dataset field, copied as-is from the source dataset.
Narrow_Eyes	Original dataset field, copied as-is from the source dataset.
No_Beard	Original dataset field, copied as-is from the source dataset.
Oval_Face	Original dataset field, copied as-is from the source dataset.
Pale_Skin	Original dataset field, copied as-is from the source dataset.
Pointy_Nose	Original dataset field, copied as-is from the source dataset.
Receding_Hairline	Original dataset field, copied as-is from the source dataset.
Rosy_Cheeks	Original dataset field, copied as-is from the source dataset.
Sideburns	Original dataset field, copied as-is from the source dataset.
Smiling	Original dataset field, copied as-is from the source dataset.
Straight_Hair	Original dataset field, copied as-is from the source dataset.
Wavy_Hair	Original dataset field, copied as-is from the source dataset.
Wearing_Earrings	Original dataset field, copied as-is from the source dataset.
Wearing_Hat	Original dataset field, copied as-is from the source dataset.
Wearing_Lipstick	Original dataset field, copied as-is from the source dataset.
Wearing_Necklace	Original dataset field, copied as-is from the source dataset.
Wearing_Necktie	Original dataset field, copied as-is from the source dataset.
Young	Original dataset field, copied as-is from the source dataset.

Available Models

Number of models with precomputed embeddings, per split:

Split	# Models
test	1699
train	1699
valid	1699

Notes

• Configs are generated from what is actually uploaded on the Hub (parquet presence).
• Based on flwrlabs/celeba
• Code: github.com/SPAICOM/semasia-datasets
• Model metadata (architecture family, parameter count, embedding dimension, pretraining details, ...) for every model in this dataset is available in the model registry.

Citation

If you use this dataset, please cite:

@misc{pandolfo2026semasialargescaledatasetsemantically,
      title={SEMASIA: A Large-Scale Dataset of Semantically Structured Latent Representations},
      author={Mario Edoardo Pandolfo and Enrico Grimaldi and Lorenzo Marinucci and Leonardo Di Nino and Simone Fiorellino and Sergio Barbarossa and Paolo Di Lorenzo},
      year={2026},
      eprint={2605.09485},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.09485},
}