Datasets:
The dataset viewer is not available for this subset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/libs/libcommon/src/libcommon/packaged_modules.py", line 18, in _refuse_lance
raise NotImplementedError(LANCE_DISABLED_MESSAGE)
NotImplementedError: The Lance format is not supported.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
DPR Wikipedia single-NQ
21,015,300 base vectors and 3,610 Natural Questions test query vectors, each with 768 float32 dimensions. All original passage vectors are retained, without normalization. The corpus contains embeddings and passage-ID mappings, not passage text.
Provenance and attribution
The passage embeddings are the original Facebook Research 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 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 distributed under CC BY-NC 4.0, also recorded in the official Wiki-DPR dataset card. The noncommercial restriction is retained. See LICENSE.
Please cite Karpukhin et al., Dense Passage Retrieval for Open-Domain Question Answering, EMNLP 2020. Credit belongs to the DPR authors and the original Wikipedia and Natural Questions contributors; this repository provides a format conversion and ANN evaluation data.
Contents
base.lance: original, unnormalized 768-dimensional float32 vectors in source order. Only thevectorcolumn is stored.queries.lance:query_id(uint32) andvector(fixed-size list of 768 float32 values).ground_truth.lance:query_id,neighbor_ids(100 uint64 values), andscores(100 values; semantics below).SHA256SUMS: checksums of the data files and manifests.
All Lance tables are self-contained and have no vector or scalar indices. The base table contains only its initial index-free manifest, with no indexed versions or index files. No source .bin, .npy, .npz, pickle shards, or benchmark indices are included.
Ground-truth neighbor IDs are zero-based row positions in the original base table, suitable for base.take(ids). They are not passage IDs or a promise about Lance physical row IDs after rewriting the table.
The base data files use Lance file format 2.2; the small auxiliary tables use 2.0. Preparation was verified with pylance 13.0.0-beta.7. Use a current reader with file format 2.2 support.
Usage
import lance
from huggingface_hub import snapshot_download
root = snapshot_download("lance-format/dpr-wikipedia-single-nq", repo_type="dataset")
base = lance.dataset(f"{root}/base.lance")
queries = lance.dataset(f"{root}/queries.lance")
ground_truth = lance.dataset(f"{root}/ground_truth.lance")
assert base.count_rows() == 21015300
assert base.list_indices() == []
query = queries.take([0]).to_pylist()[0]["vector"]
expected = ground_truth.take([0]).to_pylist()[0]["neighbor_ids"]
The benchmark metric is maximum inner product (dot in Lance). Do not normalize vectors if reproducing this benchmark. Ties at the kth boundary may admit multiple correct neighbor sets.
- Downloads last month
- 305