All datasets / lance-format

lance-format/dpr-wikipedia-single-nq

307 downloads in 30 days0 likesLast modified 21 Sep 2026View on Hugging Face

Use restrictions

The declared licence (cc-by-nc-4.0) is classified as having use restrictions.

Declared licence: cc-by-nc-4.0, from the tagEvidence: README.md at d95c371
Licence files: LICENSE, detected as cc-by-nc-4.0Evidence: LICENSE at d95c371
Declared sources: None listed in the card metadata
Access gated on Hugging Face: No

Each cell links to what was checked: a file, a file list or a Hub record.

Tasks

text-retrieval

Citation text from the card

The passage embeddings are the original [Facebook Research DPR](https://github.com/facebookresearch/DPR) single-NQ Wikipedia embeddings, downloaded in shard order 0 through 49 from `https://dl.fbaipublicfiles.com/dpr/data/wiki_encoded/single/nq/wiki_passages_{shard}`. `provenance.json` records source URLs and checksums.

Queries use the official 3,610-question NQ test split, encoded with [`facebook/dpr-question_encoder-single-nq-base`](https://huggingface.co/facebook/dpr-question_encoder-single-nq-base) at revision `d04a52f6d2f96c60117a925e8c24c4043a75f265`, in evaluation mode, with maximum token length 256 and no vector normalization. Curly apostrophes in questions were replaced with ASCII apostrophes and surrounding whitespace was stripped.

Ground truth was computed against the complete corpus using exhaustive float32 dot products to retain the best 128 candidates, followed by float64 rescoring and sorting of those candidates to publish the top 100. The `scores` column contains these **float64 inner products**, where higher is better. A 16-query audit against exhaustive float64 search was performed during benchmark preparation; this is not a claim that every query was exhaustively scored in float64. Six queries have exact ties at the top-10 boundary.

`passage_ids.lance` maps each `row_position` (uint64) to its original `passage_id`. The source shard order is preserved; passage IDs must not be substituted directly for ground-truth row positions.

The upstream DPR data is d

All datasets marked Use restrictions | JSON

The same facts for your agent

A free MCP server at mcp.datalicense.uk/mcp. Connect it from an MCP client such as Claude or Cursor.

dataset_licence({ "id": "lance-format/dpr-wikipedia-single-nq" })
signal: "restrictive"
stated_licence: "cc-by-nc-4.0"
sources: 0