neurips26/MultimodalUnlearningEvalBenchmark

🧠 Multimodal Unlearning Evaluation Benchmark

📌 Overview

This dataset provides evaluation outputs for studying metric inconsistency in multimodal machine unlearning.

It supports reproducibility of results in:

Metric Unreliability in Multimodal Machine Unlearning (NeurIPS 2026)

---

📊 Contents

File	Description
📄 multimodal_results.json	Results on VQA benchmarks (MLLMU-Bench, UnLOK-VQA, MMUBench)
📄 unimodal_results.json	CIFAR-10 baseline results
⚖️ uqs_weights.json	Learned weights for Unified Quality Score (UQS)
🏆 ranking_table.json	Method rankings across metrics
📈 analysis_results.json	Correlation and disagreement analysis
🔍 kr_pilot_results.json	Knowledge Recoverability (KR) pilot results
🤖 blip2_minimal_summary.json	Cross-architecture validation (BLIP-2)

---

🎯 Purpose

This benchmark evaluates five standard unlearning metrics:

• Forget Accuracy (FA)
• Retain Accuracy (RA)
• Membership Inference Attack (MIA)
• Activation Distance (AD)
• JS Divergence (JS)

⚠️ Key finding:

These metrics produce conflicting rankings and do not measure knowledge recoverability (KR).

---

⚙️ Usage

All results in the paper can be reproduced directly from these files.

Example:

import json

with open("multimodal_results.json") as f:
    data = json.load(f)


📚 Source Datasets (Not Included)

This benchmark builds on:

MLLMU-Bench
UnLOK-VQA
MMUBench
CIFAR-10

These datasets are not redistributed here. Please refer to their original sources.

⚖️ License

This dataset is released under the CC-BY-4.0 License.

⚠️ Notes
This dataset contains evaluation outputs, not raw training data
Designed for benchmarking and reproducibility
Prepared to support anonymous peer review
🔗 Citation
Anonymous. Metric Unreliability in Multimodal Machine Unlearning. NeurIPS 2026.

---