Duplicate-aware sparse and dense retrieval evaluation on COVID-QA: benchmark, rankings and analysis code

Benchmark, per-question retrieval rankings and analysis code for a study of sparse and dense passage retrieval on 250 expert-annotated COVID-19 questions drawn from COVID-QA (Möller et al., 2020). Contains two passage indices of 4,205 and 8,185 passages built from the same source contexts, the question sets, per-question reciprocal rank and nDCG for four retrieval systems across six sentence encoders under three relevance definitions (exact passage, duplicate-aware passage, source article), knowledge-graph pruning sweeps across 34 configurations, and all analysis notebooks. Embeddings are not included and are regenerable from the notebooks. Experiments were run on 25 and 26 September 2026.