tomaarsen/miriad-benchmark-200k

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.

This is the evaluation set. If you want to train on this domain, use tomaarsen/miriad-4.4M-split, which holds 4.4M question-passage pairs. The 9,985 gold passages here are held out from it, though the 190,015 distractors are train passages, so exclude them if you train on the full split.

Usage

from datasets import load_dataset

corpus = load_dataset("tomaarsen/miriad-benchmark-200k", "corpus", split="corpus")
queries = load_dataset("tomaarsen/miriad-benchmark-200k", "queries", split="queries")
qrels = load_dataset("tomaarsen/miriad-benchmark-200k", "qrels", split="test")

With Sentence Transformers, for a multi-vector model:

from sentence_transformers import MultiVectorEncoder
from sentence_transformers.multi_vector_encoder.evaluation import MultiVectorInformationRetrievalEvaluator

model = MultiVectorEncoder("multi-vector-encoder/mLateOn-medical")

# The headline protocol uses the first 1,000 queries
queries = queries.select(range(1000))
query_ids = set(queries["_id"])

evaluator = MultiVectorInformationRetrievalEvaluator(
    queries=dict(zip(queries["_id"], queries["text"])),
    corpus=dict(zip(corpus["_id"], corpus["text"])),
    relevant_docs={row["query-id"]: {row["corpus-id"]} for row in qrels if row["query-id"] in query_ids},
    name="miriad-200k",
)
print(evaluator(model))

InformationRetrievalEvaluator and SparseInformationRetrievalEvaluator take the same arguments for dense and sparse models.

How it was built

Everything is derived deterministically from tomaarsen/miriad-4.4M-split.

part	rows	recipe
queries	5,000	The first 5,000 questions of the eval split, in order. The headline protocol uses the first 1,000.
corpus (gold)	9,985	Every deduplicated passage of the eval split, ids p0...p9984.
corpus (distractors)	190,015	The first unique train split passages, ids d0...d190014.
qrels	5,000	One relevant passage per question, score 1.

Two details worth knowing:

• All 9,985 eval passages are in the corpus, not only the ones belonging to the sampled queries. The rest act as hard in-domain distractors.
• Distractors are deduplicated against the gold passages and against each other. A train passage identical to a gold one would be scored as a miss even when the model is right.

Passages are long, averaging 941 tokens, so check your model's document length cap before evaluating. Truncating at 180 to 512 tokens, as many released retrieval checkpoints do by default, costs up to 0.24 NDCG@10 here, more than any difference between model architectures.

Reference scores

NDCG@10 on the 1,000-query protocol, from Training and Finetuning Multi-Vector Embedding Models, which evaluated 47 model configurations across four architecture families.

model	family	NDCG@10
multi-vector-encoder/mLateOn-medical	Multi-vector, finetuned	0.9139
lightonai/mLateOn	Multi-vector	0.8520
Qwen/Qwen3-Embedding-4B	Dense	0.7817
BM25	Lexical	0.7501
naver/splade-v3	Sparse	0.6853

Citation

This dataset is derived from MIRIAD. Please cite the original work:

@misc{zheng2025miriadaugmentingllmsmillions,
      title={MIRIAD: Augmenting LLMs with millions of medical query-response pairs},
      author={Qinyue Zheng and Salman Abdullah and Sam Rawal and Cyril Zakka and Sophie Ostmeier and Maximilian Purk and Eduardo Reis and Eric J. Topol and Jure Leskovec and Michael Moor},
      year={2025},
      eprint={2506.06091},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.06091},
}