sebasmos/QuantumEmbeddings

Overview

Embedding datasets and experiment outputs for evaluating quantum-classical support vector machines (QSVMs) on features from pretrained classical encoders. Everything here supports QuantumVE, the code base of the paper:

• Code: https://github.com/sebasmos/QuantumVE
• Paper: Embedding-Aware Quantum-Classical SVMs for Scalable Quantum Machine Learning, CEUR Workshop Proceedings Vol. 4153, AIQxQIA at ECAI 2025.

QuantumVE computes embeddings with pretrained vision models (EfficientNet, ViT-B/16, ViT-B/32, ViT-L/14, ViT-L/14@336px), reduces them, and feeds them to quantum-kernel SVMs.

The data are subsets of public datasets:

• MNIST: public domain, http://yann.lecun.com/exdb/mnist/
• Fashion-MNIST: MIT License, https://github.com/zalandoresearch/fashion-mnist

Only subsets (images and embeddings) are redistributed, in accordance with their licenses, for non-commercial academic use.

Repository contents

Path	Size	What it is
mnist_embeddings.zip	77 MB	MNIST embeddings, one folder per encoder
fashionmnist_embeddings.zip	126 MB	Fashion-MNIST embeddings, one folder per encoder
experiments/EXPERIMENTS.zip	26 MB	Raw results of the QuantumVE MNIST / Fashion-MNIST SVM runs (CSV, PNG, NPZ, pickled models; 5-fold CV, 16 qubits)
experiments/README.md		Folder-by-folder index of EXPERIMENTS.zip

Inside the embedding zips, folders are named <model_name>_<embedding_dim>/ (for example efficientnet_1536/, vit_b_16_512/, vit_l_14@336px_768/), each with train.csv and test.csv. The last column of every CSV is the class label; the other columns are embedding values.

Loading

import zipfile
import pandas as pd
from huggingface_hub import hf_hub_download

p = hf_hub_download("sebasmos/QuantumEmbeddings", "mnist_embeddings.zip", repo_type="dataset")
with zipfile.ZipFile(p) as z:
    z.extractall("mnist_embeddings")

# adjust the folder to the actual layout inside the zip
df = pd.read_csv("mnist_embeddings/vit_b_16_512/train.csv")
X, y = df.iloc[:, :-1].values, df.iloc[:, -1].values

Medical work

The medical-imaging extension of this work (chest X-ray embeddings from MIMIC-CXR) is distributed separately in the gated dataset MITCriticalData/qml-mimic-cxr-embeddings, because it derives from PhysioNet data that require credentialing and a signed DUA. No medical data are in this repository.

Citation

@misc{cajas_quantumve_2025,
  author       = {Sebastián Andrés Cajas Ordóñez and Luis Torres and Mario Bifulco and Carlos Duran and Cristian Bosch and Ricardo Simon Carbajo},
  title        = {Embedding-Aware Quantum-Classical SVMs for Scalable Quantum Machine Learning},
  year         = {2025},
  howpublished = {CEUR Workshop Proceedings, Vol. 4153, AIQxQIA at ECAI 2025},
  url          = {https://github.com/sebasmos/QuantumVE}
}

License

MIT License (code and derived embeddings); see the licenses of MNIST and Fashion-MNIST above for the underlying data.

Contact

Questions or contributions: open an issue or pull request at https://github.com/sebasmos/QuantumVE.