MS MARCO (HELM passage ranking)

HELM's MS MARCO passage-ranking wrap: an LLM says Yes or No whether a passage answers a Bing query, then HELM ranks those decisions.

Also known as: MS MARCO, MSMARCO, Microsoft MAchine Reading COmprehension, msmarco_regular, msmarco_trec

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categoryknowledge
SubcategoryHELM binary LLM ranker on MS MARCO passage tracks (regular and TREC)
Page statusunknown
MetricRR@10 (regular track); NDCG@10 (TREC track)
Directionhigher_is_better
Dataset licencenon-commercial research only; no IP licence granted (MS MARCO terms)
PublisherMicrosoft (dataset); HELM wrap by Stanford CRFM

What it measures

This id is HELM's `msmarco` scenario, not the original 2016 reading- comprehension leaderboard and not an embedding nDCG run from [beir](beir.md). Microsoft MS MARCO pairs Bing queries with web passages. HELM follows Nogueira and Jiang (2020) and turns ranking into binary classification: given one passage and one query, the model must answer whether the passage answers the query. HELM then sorts passages by the Yes/No token and its log-probability. Two passage tracks exist: regular (dev.small qrels) and TREC (2019 deep-learning qrels). English text.

Task format

Binary ranking adapter. Prompt block is Passage / Query / "Does the passage answer the query?" / Answer: Yes or No. Default four in-context training instances (two blocks each: one gold, one non-gold). Stop at newline. Run spec name msmarco:track={regular|trec},valid_topk={int}. Regular gold relation is 1. TREC gold relations are 2 and 3 (0 and 1 count as non-gold).

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub