PURGE is a partitioning strategy applied to existing public datasets (MSCOCO 2017) that separates object-relevant evidence from spurious background cues in LVLMs. This repository hosts the resulting preprocessed retain/forget partitions for direct reuse.
NeurIPS 2026 Evaluations & Datasets Track, Submission #2701. Hosted under an anonymous account for double-blind review; will be transferred… See the full description on the dataset page:
https://huggingface.co/datasets/anonnnnnsub/neuripsED_2701.
