Datasets:

ArXiv:
License:
holylovenia commited on
Commit
8eaec75
1 Parent(s): 1d80dec

Upload culturax.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. culturax.py +150 -0
culturax.py ADDED
@@ -0,0 +1,150 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from pathlib import Path
2
+ from typing import Dict, List, Tuple
3
+ from urllib.parse import urljoin
4
+
5
+ import datasets
6
+ from pyarrow import parquet as pq
7
+
8
+ from seacrowd.utils import schemas
9
+ from seacrowd.utils.configs import SEACrowdConfig
10
+ from seacrowd.utils.constants import Tasks, Licenses
11
+
12
+ _CITATION = """\
13
+ @article{nguyen2023culturax,
14
+ author = {Thuat Nguyen and Chien Van Nguyen and Viet Dac Lai and Hieu Man and Nghia Trung Ngo and Franck Dernoncourt and Ryan A. Rossi and Thien Huu Nguyen},
15
+ title = {CulturaX: A Cleaned, Enormous, and Multilingual Dataset for Large Language Models in 167 Languages},
16
+ journal = {arXiv preprint arXiv:2309.09400},
17
+ year = {2023},
18
+ url = {https://arxiv.org/abs/2309.09400},
19
+ }
20
+ """
21
+
22
+ _DATASETNAME = "culturax"
23
+ _DESCRIPTION = """\
24
+ CulturaX is a comprehensive multilingual dataset comprising 6.3 trillion tokens across 167
25
+ languages, designed for large language model development. It incorporates an advanced
26
+ cleaning and deduplication process, including language identification and fuzzy
27
+ deduplication with MinHash, to ensure high-quality data for model training. The dataset,
28
+ which spans 16TB in parquet format and 27TB when unpacked, is a combination of the latest
29
+ mC4 and OSCAR corpora, emphasizing non-English languages to support multilingual model
30
+ training. For data cleaning validation, CulturaX employs a SentencePiece tokenizer and
31
+ KenLM language models, utilizing recent Wikipedia dumps for perplexity scoring.
32
+ Before using this dataloader, please accept the acknowledgement at https://huggingface.co/datasets/uonlp/CulturaX and use huggingface-cli login for authentication.
33
+ """
34
+
35
+ _LOCAL=False
36
+ _LANGUAGES = ["ind", "jav", "khm", "lao", "tgl", "min", "mya", "sun", "tha", "vie", "zlm", "ceb", "war", "cbk", "bcl"] # We follow ISO639-3 language code (https://iso639-3.sil.org/code_tables/639/data)
37
+
38
+ _HOMEPAGE = "https://huggingface.co/datasets/uonlp/CulturaX"
39
+ _LICENSE = f"""{Licenses.OTHERS.value} | \
40
+ The licence terms for CulturaX strictly follows those of mC4 and OSCAR. \
41
+ Please refer to both below licenses when using this dataset. \
42
+ - mC4 license: https://huggingface.co/datasets/allenai/c4#license \
43
+ - OSCAR license: https://huggingface.co/datasets/oscar-corpus/OSCAR-2301#licensing-information \
44
+ """
45
+ _BASE_URL = "https://huggingface.co/datasets/uonlp/CulturaX/resolve/main/{lang}/"
46
+
47
+ _SUPPORTED_TASKS = [Tasks.SELF_SUPERVISED_PRETRAINING]
48
+ _SOURCE_VERSION = "1.0.0"
49
+ _SEACROWD_VERSION = "2024.06.20"
50
+
51
+
52
+ class CulturaXDataset(datasets.GeneratorBasedBuilder):
53
+ """CulturaX subset for SEA languages."""
54
+
55
+ SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
56
+ SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION)
57
+
58
+ SEACROWD_SCHEMA_NAME = "ssp"
59
+ SUBSETS = ["id", "jv", "km", "lo", "tl", "min", "my", "su", "th", "vi", "ms", "ceb", "war", "cbk", "bcl"]
60
+
61
+ BUILDER_CONFIGS = [
62
+ SEACrowdConfig(
63
+ name=f"{_DATASETNAME}_{subset}_source",
64
+ version=datasets.Version(_SOURCE_VERSION),
65
+ description=f"{_DATASETNAME} {subset} source schema",
66
+ schema="source",
67
+ subset_id=subset,
68
+ )
69
+ for subset in SUBSETS
70
+ ] + [
71
+ SEACrowdConfig(
72
+ name=f"{_DATASETNAME}_{subset}_seacrowd_ssp",
73
+ version=datasets.Version(_SEACROWD_VERSION),
74
+ description=f"{_DATASETNAME} {subset} SEACrowd schema",
75
+ schema="seacrowd_ssp",
76
+ subset_id=subset,
77
+ )
78
+ for subset in SUBSETS
79
+ ]
80
+
81
+ DEFAULT_CONFIG_NAME = f"{_DATASETNAME}_jv_source"
82
+
83
+ def _info(self) -> datasets.DatasetInfo:
84
+ if self.config.schema == "source":
85
+ features = datasets.Features(
86
+ {
87
+ "text": datasets.Value("string"),
88
+ "timestamp": datasets.Value("string"),
89
+ "url": datasets.Value("string"),
90
+ "source": datasets.Value("string"),
91
+ }
92
+ )
93
+ elif self.config.schema == f"seacrowd_{self.SEACROWD_SCHEMA_NAME}":
94
+ features = schemas.ssp_features
95
+
96
+ return datasets.DatasetInfo(
97
+ description=_DESCRIPTION,
98
+ features=features,
99
+ homepage=_HOMEPAGE,
100
+ license=_LICENSE,
101
+ citation=_CITATION,
102
+ )
103
+
104
+ def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
105
+ """Returns SplitGenerators."""
106
+ base_path = _BASE_URL.format(lang=self.config.name.split("_")[1])
107
+
108
+ checksum_url = urljoin(base_path, "checksum.sha256")
109
+ checksum_path = Path(dl_manager.download(checksum_url))
110
+ with open(checksum_path, encoding="utf-8") as f:
111
+ filenames = [line.split()[1] for line in f if line]
112
+ data_urls = [urljoin(base_path, filename) for filename in filenames]
113
+
114
+ data_paths = list(map(Path, dl_manager.download([url for url in data_urls if url.endswith(".parquet")])))
115
+
116
+ return [
117
+ datasets.SplitGenerator(
118
+ name=datasets.Split.TRAIN,
119
+ gen_kwargs={
120
+ "filepaths": data_paths,
121
+ "split": "train",
122
+ },
123
+ )
124
+ ]
125
+
126
+ def _generate_examples(self, filepaths: [Path], split: str) -> Tuple[int, Dict]:
127
+ """Yields examples as (key, example) tuples.
128
+
129
+ Iterate over row groups in each filepaths, then yield each row as an example.
130
+ """
131
+ key = 0
132
+ for filepath in filepaths:
133
+ with open(filepath, "rb") as f:
134
+ pf = pq.ParquetFile(f)
135
+ for row_group in range(pf.num_row_groups):
136
+ df = pf.read_row_group(row_group).to_pandas()
137
+ for row in df.itertuples():
138
+ if self.config.schema == "source":
139
+ yield key, {
140
+ "text": row.text,
141
+ "timestamp": row.timestamp,
142
+ "url": row.url,
143
+ "source": row.source,
144
+ }
145
+ elif self.config.schema == f"seacrowd_{self.SEACROWD_SCHEMA_NAME}":
146
+ yield key, {
147
+ "id": str(key),
148
+ "text": row.text,
149
+ }
150
+ key += 1