microsoft / msmarco-document-ranking Goto Github PK

Hey,
I am confused about the meaning of "offset_trec" and "offset_tsv" in corpus msmarco-docs-lookup.tsv , how to use this corpus to get a train dataset having format like "qid \t docid \t label"?

Hello,
I had hard time to understand dataset structure and I was not able to find it anywhere.

I guess that:

• fulldocs.tsv are urls and documents and there are 3213835 lines (so each line is one document)
• docleaderboard-queries.tsv are ids and queries.
• docleaderboard-top100.tsv are query to document relevancy dataset.
  Structure:

355339 Q0 D2612180 1 -5.21639 IndriQueryLikelihood

355339 Q0 D2906504 2 -5.22419 IndriQueryLikelihood

355339 Q0 D2076296 3 -5.39244 IndriQueryLikelihood

355339 Q0 D894964 4 -5.46692 IndriQueryLikelihood

355339 Q0 D260320 5 -5.47119 IndriQueryLikelihood

I assume first column is question ID, the second I don't know, the third possibly document id and next row is relevancy from 1-10, rest I don't care.
If DXXXXXX is document ID and as long as fulldocs.tsv has no ID in row, then I assume it is row number.
But if this is true, maximum DXXXXX number from docleaderboard-top100.tsv is D3563531 which is more than lines in fulldocs.tsv.

I am lost. Can you please help me. Is it possible that fulldocs.tsv is missing some docs ?

Thank you

Hi, I am trying to download the ms-marco dataset from here.
However, it doesn't matter what version I want, I get an error: 'ResourceNotFound'.

Is there a mirror link to the dataset's file?

1. It seems like test queries point to TREC DL 19 queries? Is this correct and if so I feel like the records count needs to be updated to 200?
2. I feel like the records count for the dev set needs to be 5,193?