Data for "Architecture Matters: Comparing RAG Systems under Knowledge Base Poisoning"

Experiment results, LLM judge classifications, noise filter outputs, human validation labels, and LLM-generated poisoned documents for the paper "Architecture Matters: Comparing RAG Systems under Knowledge Base Poisoning." Contains results from 12 experiments (4 RAG architectures × 3 attack conditions) on 921 gold-filtered questions from Natural Questions, plus LLM judge outputs, noise filtering results, and 1,475 manually labeled validation instances. Code and analysis notebooks: https://github.com/samkorn/rag-poisoning-architecture-bench