Datasets:

Languages:
English
Multilinguality:
multilingual
Size Categories:
10M<n<100M
Language Creators:
crowdsourced
Annotations Creators:
no-annotation
Source Datasets:
original
ArXiv:
License:
File size: 9,930 Bytes
80d4f4b
 
 
 
 
 
 
3261af0
 
 
80d4f4b
3686636
 
 
 
 
 
 
 
 
 
 
80d4f4b
 
 
 
 
 
 
 
 
 
 
 
 
cdb0294
 
 
80d4f4b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cdb0294
80d4f4b
cdb0294
 
 
 
 
80d4f4b
 
 
3261af0
80d4f4b
cdb0294
80d4f4b
 
 
 
3261af0
 
80d4f4b
 
 
 
 
cdb0294
 
 
80d4f4b
 
 
cdb0294
80d4f4b
cdb0294
80d4f4b
cdb0294
80d4f4b
 
 
 
 
 
cdb0294
 
 
3261af0
cdb0294
 
 
3261af0
cdb0294
 
 
80d4f4b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cdb0294
80d4f4b
cdb0294
80d4f4b
cdb0294
80d4f4b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e15af35
80d4f4b
 
 
 
e15af35
80d4f4b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cdb0294
80d4f4b
 
f775f65
80d4f4b
 
 
 
 
 
 
 
 
 
 
 
cdb0294
80d4f4b
3261af0
80d4f4b
3261af0
cdb0294
 
3261af0
80d4f4b
cdb0294
 
 
 
 
80d4f4b
 
 
 
 
 
 
3261af0
80d4f4b
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
import os

import numpy as np

import datasets


logger = datasets.logging.get_logger(__name__)


_CITATION = """
@inproceedings{karpukhin-etal-2020-dense,
    title = "Dense Passage Retrieval for Open-Domain Question Answering",
    author = "Karpukhin, Vladimir and Oguz, Barlas and Min, Sewon and Lewis, Patrick and Wu, Ledell and Edunov, Sergey and Chen, Danqi and Yih, Wen-tau",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.550",
    doi = "10.18653/v1/2020.emnlp-main.550",
    pages = "6769--6781",
}
"""

_DESCRIPTION = """
This is the wikipedia split used to evaluate the Dense Passage Retrieval (DPR) model.
It 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.
"""

_LICENSE = """DPR is CC-BY-NC 4.0 licensed."""

_DATA_URL = "https://dl.fbaipublicfiles.com/dpr/wikipedia_split/psgs_w100.tsv.gz"

_NQ_VECTORS_URL = "https://dl.fbaipublicfiles.com/dpr/data/wiki_encoded/single/nq/wiki_passages_{i}"

_MULTISET_VECTORS_URL = "https://dl.fbaipublicfiles.com/rag/rag_multiset_embeddings/wiki_passages_{i}"

_INDEX_URL = "https://storage.googleapis.com/huggingface-nlp/datasets/wiki_dpr"


class WikiDprConfig(datasets.BuilderConfig):
    """BuilderConfig for WikiDprConfig."""

    def __init__(
        self,
        with_embeddings=True,
        with_index=True,
        wiki_split="psgs_w100",
        embeddings_name="nq",
        index_name="compressed",
        index_train_size=262144,
        dummy=False,
        **kwargs,
    ):
        """BuilderConfig for WikiSnippets.
        Args:
            with_embeddings (`bool`, defaults to `True`): Load the 768-dimensional embeddings from DPR.
            with_index (`bool`, defaults to `True`): Load the faiss index trained on the embeddings.
            wiki_split (`str`, defaults to `psgs_w100`): name of the splitting method of wiki articles.
            embeddings_name (`str`, defaults to `nq`): "nq" or "multiset", depending on which dataset DPR was trained on.
            index_name (`str`, defaults to `compressed`): "compressed" or "exact", the configuration of the faiss index to use.
            index_train_size (`int`, defaults to `262144`): Size of the subset to use to train the index, if it is trainable.
            dummy (`bool`, defaults to `False`): Dummy uses only 10 000 examples for testing purposes.
          **kwargs: keyword arguments forwarded to super.
        """
        self.with_embeddings = with_embeddings
        self.with_index = with_index and index_name != "no_index"
        self.wiki_split = wiki_split
        self.embeddings_name = embeddings_name
        self.index_name = index_name if with_index else "no_index"
        self.index_train_size = index_train_size
        self.dummy = dummy
        name = [self.wiki_split, self.embeddings_name, self.index_name]
        if not self.with_embeddings:
            name.append("no_embeddings")
        if self.dummy:
            name = ["dummy"] + name
            assert (
                self.index_name != "compressed" or not self.with_index
            ), "Please use `index_name='exact' for dummy wiki_dpr`"
        assert wiki_split == "psgs_w100"
        assert embeddings_name in ("nq", "multiset")
        assert index_name in ("compressed", "exact", "no_index")
        kwargs["name"] = ".".join(name)
        super(WikiDprConfig, self).__init__(**kwargs)

        prefix = f"{wiki_split}.{embeddings_name}."
        if self.index_name == "exact":
            self.index_file = prefix + "HNSW128_SQ8-IP-{split}.faiss"
        else:
            self.index_file = prefix + "IVF4096_HNSW128_PQ128-IP-{split}.faiss"
        if self.dummy:
            self.index_file = "dummy." + self.index_file


class WikiDpr(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIG_CLASS = WikiDprConfig
    BUILDER_CONFIGS = [
        WikiDprConfig(
            embeddings_name=embeddings_name,
            with_embeddings=with_embeddings,
            index_name=index_name,
            version=datasets.Version("0.0.0"),
        )
        for with_embeddings in (True, False)
        for embeddings_name in ("nq", "multiset")
        for index_name in ("exact", "compressed", "no_index")
    ]

    def _info(self):
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=datasets.Features(
                {
                    "id": datasets.Value("string"),
                    "text": datasets.Value("string"),
                    "title": datasets.Value("string"),
                    "embeddings": datasets.Sequence(datasets.Value("float32")),
                }
            )
            if self.config.with_embeddings
            else datasets.Features(
                {"id": datasets.Value("string"), "text": datasets.Value("string"), "title": datasets.Value("string")}
            ),
            supervised_keys=None,
            homepage="https://github.com/facebookresearch/DPR",
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        files_to_download = {"data_file": _DATA_URL}
        downloaded_files = dl_manager.download_and_extract(files_to_download)
        if self.config.with_embeddings:
            vectors_url = _NQ_VECTORS_URL if self.config.embeddings_name == "nq" else _MULTISET_VECTORS_URL
            if self.config.dummy:
                downloaded_files["vectors_files"] = dl_manager.download([vectors_url.format(i=0)])
            else:
                downloaded_files["vectors_files"] = dl_manager.download([vectors_url.format(i=i) for i in range(50)])
        return [
            datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs=downloaded_files),
        ]

    def _generate_examples(self, data_file, vectors_files=None):
        vec_idx = 0
        vecs = []
        lines = open(data_file, "r", encoding="utf-8")
        next(lines)  # skip headers
        for i, line in enumerate(lines):
            if self.config.dummy and i == 10000:
                break
            if i == 21015300:
                break  # ignore the last 24 examples for which the embeddings are missing.
            id, text, title = line.strip().split("\t")
            text = text[1:-1]  # remove " symbol at the beginning and the end
            text = text.replace('""', '"')  # replace double quotes by simple quotes
            if self.config.with_embeddings:
                if vec_idx >= len(vecs):
                    if len(vectors_files) == 0:
                        logger.warning(f"Ran out of vector files at index {i}")
                        break
                    vecs = np.load(open(vectors_files.pop(0), "rb"), allow_pickle=True)
                    vec_idx = 0
                vec_id, vec = vecs[vec_idx]
                assert int(id) == int(vec_id), f"ID mismatch between lines {id} and vector {vec_id}"
                yield id, {"id": id, "text": text, "title": title, "embeddings": vec}
                vec_idx += 1
            else:
                yield id, {
                    "id": id,
                    "text": text,
                    "title": title,
                }

    def _post_processing_resources(self, split):
        if self.config.with_index:
            return {"embeddings_index": self.config.index_file.format(split=split)}
        else:
            return {}

    def _download_post_processing_resources(self, split, resource_name, dl_manager):
        if resource_name == "embeddings_index":
            try:
                downloaded_resources = dl_manager.download_and_extract(
                    {"embeddings_index": _INDEX_URL + "/" + self.config.index_file.format(split=split)}
                )
                return downloaded_resources["embeddings_index"]
            except (FileNotFoundError, ConnectionError):  # index doesn't exist
                pass

    def _post_process(self, dataset, resources_paths):
        if self.config.with_index:
            index_file = resources_paths["embeddings_index"]
            if os.path.exists(index_file):
                dataset.load_faiss_index("embeddings", index_file)
            else:
                if "embeddings" not in dataset.column_names:
                    raise ValueError("Couldn't build the index because there are no embeddings.")
                import faiss

                d = 768
                train_size = self.config.index_train_size
                logger.info("Building wiki_dpr faiss index")
                if self.config.index_name == "exact":
                    index = faiss.IndexHNSWSQ(d, faiss.ScalarQuantizer.QT_8bit, 128, faiss.METRIC_INNER_PRODUCT)
                    index.hnsw.efConstruction = 200
                    index.hnsw.efSearch = 128
                    dataset.add_faiss_index("embeddings", custom_index=index, train_size=train_size)
                else:
                    quantizer = faiss.IndexHNSWFlat(d, 128, faiss.METRIC_INNER_PRODUCT)
                    quantizer.hnsw.efConstruction = 200
                    quantizer.hnsw.efSearch = 128
                    ivf_index = faiss.IndexIVFPQ(quantizer, d, 4096, 128, 8, faiss.METRIC_INNER_PRODUCT)
                    ivf_index.nprobe = 64
                    ivf_index.own_fields = True
                    quantizer.this.disown()
                    dataset.add_faiss_index(
                        "embeddings",
                        train_size=train_size,
                        custom_index=ivf_index,
                    )
                logger.info("Saving wiki_dpr faiss index")
                dataset.save_faiss_index("embeddings", index_file)
        return dataset