Datasets:
metadata
license: apache-2.0
language:
- en
pretty_name: Wiki DPR - Lance example
Wiki-DPR (first 10k samples of the "multiset" version)
Overview
This is the wikipedia split used to evaluate the Dense Passage Retrieval (DPR) model. The original dataset contains 21M passages from wikipedia along with their DPR embeddings. The wikipedia articles were split into multiple, disjoint text blocks of 100 words as passages.
Embeddings
Passages are embedded using this model: https://huggingface.co/sentence-transformers/facebook-dpr-ctx_encoder-multiset-base
Questions should embedded using this model: https://huggingface.co/sentence-transformers/facebook-dpr-question_encoder-multiset-base
Dataset format
The dataset is in Lance format and contains an index over the embeddings column, created using
>>> table.create_index(metric="dot", vector_column_name="embeddings")
Usage example
>>> import lance
>>> from sentence_transformers import SentenceTransformer
>>> ds = lance.dataset("hf://datasets/lhoestq/wiki-dpr-lance-example/dummy-multiset.lance")
>>> ds.has_index
True
>>> q = embeddings = model.encode(["In what year did ABBA start ?"])[0]
>>> df = ds.to_table(nearest={"column": "embeddings", "q": embeddings, "k": 5}).to_pandas()
>>> df
embeddings id text title _distance
0 [-0.18065824, 0.25305566, 0.1199052, -0.571353... 9929 numbers from respective albums, but the lukewa... ABBA -71.944893
1 [-0.40663224, 0.22715698, 0.06069667, -0.33584... 9934 small label like Playboy Records did not have ... ABBA -71.606918
2 [-0.11441193, -0.08622038, 0.1414486, -0.53792... 9931 into the mainstream international market with ... ABBA -71.264702
3 [-0.10480262, 0.24277505, 0.0049878294, -0.406... 9930 on her own. Frequent recording sessions brough... ABBA -71.059311
4 [-0.013585132, 0.010848946, 0.29921767, -0.283... 9933 credited to Björn & Benny, Agnetha & Anni-Fri... ABBA -71.030861
>>> df.iloc[0].text[435:544]
"It was during 1971 that the four artists began working together more, adding vocals to the others' recordings"