Datasets:
ArXiv:
License:
holylovenia
commited on
Commit
•
994dad2
1
Parent(s):
d0607a7
Upload culturay.py with huggingface_hub
Browse files- culturay.py +177 -0
culturay.py
ADDED
@@ -0,0 +1,177 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2022 The HuggingFace Datasets Authors and the current dataset script contributor.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
|
16 |
+
import io
|
17 |
+
import json
|
18 |
+
from pathlib import Path
|
19 |
+
from typing import Dict, List, Tuple
|
20 |
+
|
21 |
+
import datasets
|
22 |
+
import zstandard as zstd
|
23 |
+
from huggingface_hub import HfFileSystem
|
24 |
+
|
25 |
+
from seacrowd.utils.configs import SEACrowdConfig
|
26 |
+
from seacrowd.utils.constants import (SCHEMA_TO_FEATURES, TASK_TO_SCHEMA,
|
27 |
+
Licenses, Tasks)
|
28 |
+
|
29 |
+
_CITATION = """\
|
30 |
+
@misc{nguyen2024culturay,
|
31 |
+
title={CulturaY: A Large Cleaned Multilingual Dataset of 75 Languages},
|
32 |
+
author={Thuat Nguyen, Huu Nguyen and Thien Nguyen},
|
33 |
+
year={2024},
|
34 |
+
}
|
35 |
+
"""
|
36 |
+
|
37 |
+
_DATASETNAME = "culturay"
|
38 |
+
|
39 |
+
_DESCRIPTION = """\
|
40 |
+
CulturaY: A Large Cleaned Multilingual Dataset of 75 Languages From the team
|
41 |
+
that brought you CulutraX, we present CulturaY, another substantial multilingual
|
42 |
+
dataset of 15TB (uncompressed)/3TB (zstd-compressed) that applies the same
|
43 |
+
dataset cleaning methodology to the HPLT v1.1 dataset. Please note that HPLT
|
44 |
+
v1.2 has also been released and is an alternative verison with different
|
45 |
+
cleaning methodolgies. This data was used in part to train our SOTA Vietnamese
|
46 |
+
model: Vistral-7B-Chat.
|
47 |
+
|
48 |
+
Before using this dataloader, please accept the acknowledgement at
|
49 |
+
https://huggingface.co/datasets/ontocord/CulturaY and use huggingface-cli login
|
50 |
+
for authentication.
|
51 |
+
"""
|
52 |
+
|
53 |
+
_HOMEPAGE = "https://huggingface.co/datasets/ontocord/CulturaY"
|
54 |
+
|
55 |
+
_LANGUAGES = ["mya", "fil", "zlm", "vie", "ind", "tha"]
|
56 |
+
|
57 |
+
_LICENSE = Licenses.CC_BY_4_0.value
|
58 |
+
|
59 |
+
_LOCAL = False
|
60 |
+
|
61 |
+
_BASE_URL = "https://huggingface.co/datasets/ontocord/CulturaY/resolve/main/{lang}/"
|
62 |
+
|
63 |
+
_SUPPORTED_TASKS = [Tasks.SELF_SUPERVISED_PRETRAINING]
|
64 |
+
_SEACROWD_SCHEMA = f"seacrowd_{TASK_TO_SCHEMA[_SUPPORTED_TASKS[0]].lower()}" # ssp
|
65 |
+
|
66 |
+
_SOURCE_VERSION = "1.0.0"
|
67 |
+
|
68 |
+
_SEACROWD_VERSION = "2024.06.20"
|
69 |
+
|
70 |
+
|
71 |
+
class CulturaYDataset(datasets.GeneratorBasedBuilder):
|
72 |
+
"""Substantial multilingual dataset by cleaning HPLT v1.1 (Internet Archive) data."""
|
73 |
+
|
74 |
+
SOURCE_VERSION = datasets.Version(_SOURCE_VERSION)
|
75 |
+
SEACROWD_VERSION = datasets.Version(_SEACROWD_VERSION)
|
76 |
+
|
77 |
+
BUILDER_CONFIGS = []
|
78 |
+
for subset in _LANGUAGES:
|
79 |
+
BUILDER_CONFIGS += [
|
80 |
+
SEACrowdConfig(
|
81 |
+
name=f"{_DATASETNAME}_{subset}_source",
|
82 |
+
version=SOURCE_VERSION,
|
83 |
+
description=f"{_DATASETNAME} {subset} source schema",
|
84 |
+
schema="source",
|
85 |
+
subset_id=subset,
|
86 |
+
),
|
87 |
+
SEACrowdConfig(
|
88 |
+
name=f"{_DATASETNAME}_{subset}_{_SEACROWD_SCHEMA}",
|
89 |
+
version=SEACROWD_VERSION,
|
90 |
+
description=f"{_DATASETNAME} {subset} SEACrowd schema",
|
91 |
+
schema=_SEACROWD_SCHEMA,
|
92 |
+
subset_id=subset,
|
93 |
+
),
|
94 |
+
]
|
95 |
+
|
96 |
+
DEFAULT_CONFIG_NAME = f"{_DATASETNAME}_my_source" # smallest wrt n_doc
|
97 |
+
|
98 |
+
def _info(self) -> datasets.DatasetInfo:
|
99 |
+
if self.config.schema == "source":
|
100 |
+
features = datasets.Features(
|
101 |
+
{
|
102 |
+
"id": datasets.Value("int64"),
|
103 |
+
"document_lang": datasets.Value("string"),
|
104 |
+
"scores": datasets.Sequence(datasets.Value("float64")),
|
105 |
+
"langs": datasets.Sequence(datasets.Value("string")),
|
106 |
+
"text": datasets.Value("string"),
|
107 |
+
"url": datasets.Value("string"),
|
108 |
+
"collection": datasets.Value("string"),
|
109 |
+
}
|
110 |
+
)
|
111 |
+
elif self.config.schema == _SEACROWD_SCHEMA:
|
112 |
+
features = SCHEMA_TO_FEATURES[
|
113 |
+
TASK_TO_SCHEMA[_SUPPORTED_TASKS[0]]
|
114 |
+
] # ssp_features
|
115 |
+
|
116 |
+
return datasets.DatasetInfo(
|
117 |
+
description=_DESCRIPTION,
|
118 |
+
features=features,
|
119 |
+
homepage=_HOMEPAGE,
|
120 |
+
license=_LICENSE,
|
121 |
+
citation=_CITATION,
|
122 |
+
)
|
123 |
+
|
124 |
+
def _split_generators(self, dl_manager: datasets.DownloadManager) -> List[datasets.SplitGenerator]:
|
125 |
+
"""Returns SplitGenerators. Data is not yet extracted for efficient generation."""
|
126 |
+
lang_dict = {"mya": "my", "fil": "tl", "zlm": "ms", "vie": "vi", "ind": "id", "tha": "th"}
|
127 |
+
subset = lang_dict[self.config.subset_id]
|
128 |
+
base_path = _BASE_URL.format(lang=subset)
|
129 |
+
|
130 |
+
fs = HfFileSystem(token=dl_manager.download_config.token)
|
131 |
+
file_list = fs.ls(f"datasets/ontocord/CulturaY/{subset}", detail=False)
|
132 |
+
|
133 |
+
data_urls = [
|
134 |
+
f"{base_path}{filename.split('/')[-1]}"
|
135 |
+
for filename in file_list
|
136 |
+
if filename.endswith(".jsonl.zst")
|
137 |
+
]
|
138 |
+
|
139 |
+
data_paths = list(map(Path, dl_manager.download(data_urls)))
|
140 |
+
return [
|
141 |
+
datasets.SplitGenerator(
|
142 |
+
name=datasets.Split.TRAIN,
|
143 |
+
gen_kwargs={
|
144 |
+
"data_paths": data_paths,
|
145 |
+
},
|
146 |
+
),
|
147 |
+
]
|
148 |
+
|
149 |
+
def _generate_examples(self, data_paths: Path) -> Tuple[int, Dict]:
|
150 |
+
"""Yields examples as (key, example) tuples."""
|
151 |
+
key = 0
|
152 |
+
for data_path in data_paths:
|
153 |
+
with open(data_path, "rb") as f:
|
154 |
+
# Zstandard decompression
|
155 |
+
dctx = zstd.ZstdDecompressor()
|
156 |
+
reader = dctx.stream_reader(f)
|
157 |
+
text_io = io.TextIOWrapper(reader, encoding="utf-8")
|
158 |
+
|
159 |
+
# read jsonl file by line and yield
|
160 |
+
for line in text_io:
|
161 |
+
data = json.loads(line)
|
162 |
+
if self.config.schema == "source":
|
163 |
+
yield key, {
|
164 |
+
"id": data["id"],
|
165 |
+
"document_lang": data["document_lang"],
|
166 |
+
"scores": data["scores"],
|
167 |
+
"langs": data["langs"],
|
168 |
+
"text": data["text"],
|
169 |
+
"url": data["url"],
|
170 |
+
"collection": data["collection"],
|
171 |
+
}
|
172 |
+
elif self.config.schema == _SEACROWD_SCHEMA:
|
173 |
+
yield key, {
|
174 |
+
"id": str(data["id"]),
|
175 |
+
"text": data["text"],
|
176 |
+
}
|
177 |
+
key += 1
|