miriad-benchmark-200k

简介

MIRIAD Benchmark 200k A medical retrieval benchmark: 1,000 questions searching 200,000 passages, in the BEIR layout (corpus, queries, qrels) so it works with standard IR evaluators out of the box. It is deliberately hard. Scoring questions against only their own ~10k source passages saturates above 0.97 NDCG@10 for almost every model, which tells you nothing. Adding 190,015 deduplicated distractor passages spreads the field out across roughly 0.43 to 0.91. [!TIP] This is the… See the full description on the dataset page: https://huggingface.co/datasets/tomaarsen/miriad-benchmark-200k.