DUSK: Benchmark for Machine Unlearning in Multi-Source Settings

DUSK is a benchmark dataset designed for evaluating machine unlearning in multi-source settings, where specific data sources must be forgotten while preserving others. Created by AI-ISL and last updated on 2025-05-16, it challenges algorithms to precisely erase only what must be forgotten while preserving knowledge supported by other sources.

Use Cases

• Evaluating unlearning algorithm precision based on the multi-source setting described
• Testing knowledge retention after targeted forgetting based on the factual overlap scenario
• Benchmarking model performance on tasks requiring selective memory erasure

Strengths

• Designed for a specific, advanced ML task: evaluating machine unlearning in multi-source settings
• Addresses a realistic scenario where documents share factual overlap with public content

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment

Provenance

Source

AI-ISL

Freshness

Last updated 2025-05-16 06:03:54