cvc-mmu/salmubench-512-base

SALMUBench Base Dataset

Original SALMUBench dataset before redistribution of samples.

Paper: "SALMUBench: A Benchmark for Sensitive Association-Level Multimodal Unlearning" (CVPR 2026)

---

Dataset description

This dataset corresponds to the original version of the SALMUBench benchmark dataset before redistribution of samples used in the official evaluation protocol.

Images are resized so that the largest dimension is 512 pixels.

---

Purpose

This dataset is provided to allow researchers to:

• create alternative benchmark splits
• design new evaluation protocols
• reproduce or extend SALMUBench experiments

For the official benchmark evaluation, the redistributed dataset should be used.

---

Related artifacts

Official SALMUBench evaluation dataset (redistributed): salmubench-512-redistributed

Official SALMU training dataset (redistributed): salmu-512-redistributed

Corresponding training dataset (base; without redistribution): salmu-512-base

Project repository: SALMUBench GitHub repository

---

Data Disclaimer

All identities and attributes in this dataset are synthetically generated. Any resemblance to real persons or real-world data (e.g., phone numbers) is purely coincidental.

The dataset is intended solely for research on machine unlearning and privacy-preserving machine learning.

---

Citation

@misc{selvassala2026salmubenchbenchmarksensitiveassociationlevel,
      title={SALMUBench: A Benchmark for Sensitive Association-Level Multimodal Unlearning}, 
      author={Cai Selvas-Sala and Lei Kang and Lluis Gomez},
      year={2026},
      eprint={2603.26316},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.26316}, 
}