Datasets:

---

mteb

/

arguana

like 1

Follow

Massive Text Embedding Benchmark 627

According to the paper by Boteva et al. (2016) the Arguana dataset provides

relevance links at 3 levels for thousands of queries in layman’s English to documents consisting of abstracts of research article

which is detailed as follows:

We assign relevance levels to a query q with respect to a scientific document d from three possible values: The most relevant level (2) corresponds to a direct link from q to d from the cited sources section of a page, the next level (1) is used if there exists another query q′ that directly links to d and also q’s text contains an internal link to q′. Finally, the lowest level of (0) is reserved for every marginally relevant q′ and document d.

However, the HF dataset only contains binary labels. Is this a simplified version or why are the 3 levels of relevance not included?

Thanks,

Manuel