chameleon-sentence-bert-irrelevant-retrieval

Summary

For distractor experiments, the "irrelevant" passages the system must ignore are picked by measuring semantic similarity with a specific Sentence-BERT embedding model (multi-qa-mpnet-base-dot-v1), so the difficulty and character of those distractors are tied to how that particular model judges relevance. This matters because any conclusion about the system's robustness to noise depends on whether that embedding model actually surfaces truly off-topic passages or just ones that happen to score low under one embedding space.