spaicom-lab/semasia-imagenet-1k

Latents for imagenet-1k (timm)

<h5 align="center">

&nbsp;&nbsp;&nbsp;

<br>

</h5>

This repository hosts precomputed latent representations (embeddings) extracted from timm image-classification backbones on imagenet-1k, 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-imagenet-1k",  # 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)
label = torch.tensor(ds["label"])  # (N,)

Or read the Parquet files directly with polars:

import polars as pl

df = pl.read_parquet(
    "hf://datasets/spaicom-lab/semasia-imagenet-1k/test/aimv2_1b_patch14_224.apple_pt/*.parquet"
)

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

Fields

Columns available in each Parquet file for this dataset:

Available Models

Number of models with precomputed embeddings, per split:

Notes

• —Configs are generated from what is actually uploaded on the Hub (parquet presence).
• —Based on ILSVRC/imagenet-1k
• —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},
}