sebasmos/QuantumEmbeddings

Overview

This repository provides a collection of embedding datasets for evaluating quantum-classical support vector machines (QSVMs) using embeddings from pre-trained classical models. Each dataset follows the naming convention:

<model_name>_<embedding_dim>.csv

Where:

• model_name: the architecture used to generate the embeddings (e.g., vit_b_16, efficientnet, vit_l_14@336px)
• embedding_dim: the dimensionality of the embedding vectors
• The last column in each CSV represents the class label

Each CSV file is structured as a table:

• Rows correspond to distilled training or testing samples
• Columns represent embedding values followed by the class label

These datasets are designed to support quantum kernel methods and hybrid pipelines, as implemented in QuantumVE.

This dataset collection includes subsets derived from public datasets:

• MNIST – in the public domain and freely available here
• Fashion-MNIST – distributed under the MIT License

We confirm that only subsets of these datasets are included (images and embeddings), and they are redistributed in accordance with their respective licenses. This repository is intended for non-commercial academic use.

---

Repository Structure

Datasets are stored in .csv format and organized into folders named according to the embedding configuration:

<model_name>_<embedding_dim>/
├── train.csv   # Training set with embeddings and labels
└── test.csv    # Testing set with embeddings and labels

Example Folders:

• efficientnet_1536/
• vit_b_16_512/
• vit_l_14@336px_768/

---

Loading Datasets

Example code to load and preview a dataset:

import pandas as pd

df = pd.read_csv("vit_b_16_512/train.csv")
X = df.iloc[:, :-1].values  # Embedding features
y = df.iloc[:, -1].values   # Class labels

---

Citation

If you use this dataset collection, please cite:

@misc{cajas_quantumve_2025,
  author       = {Sebastián Andrés Cajas Ordóñez, Luis Torres, Mario Bifulco, Carlos Duran, Cristian Bosch, Ricardo Simon Carbajo},
  title        = {Embedding-Aware Quantum-Classical SVMs for Scalable Quantum Machine Learning},
  year         = {2025},
  url          = {https://github.com/sebasmos/QuantumVE},
  note         = {GitHub repository},
  version      = {v1.0},
  howpublished = {\url{https://github.com/sebasmos/QuantumVE}},

---

License

QuantumVE is free and open source, released under the MIT License.

---

Contact & Contributions

This dataset collection is maintained as part of the QuantumVE project.

For questions or contributions, feel free to reach out or submit a pull request via the GitHub repository.