Rushi2002/

poisoned-context-testbed

	Poisoned Context Testbed




	
	Dataset Description

This dataset is part of the research work "RW-Steering: Rescorla-Wagner Steering of LLMs for Undesired Behaviors over Disproportionate Inappropriate Context" (EMNLP 2025). It provides a comprehensive testbed for studying LLM robustness when helpful context is mixed with inappropriate content.

The dataset contains poisoned context scenarios where legitimate information is combined with inappropriate content… See the full description on the dataset page: https://huggingface.co/datasets/Rushi2002/poisoned-context-testbed.