|
import json |
|
from dataclasses import dataclass |
|
from string import Template |
|
|
|
import datasets |
|
from datasets.download.download_manager import DownloadManager |
|
|
|
_CITATION = "" |
|
_DESCRIPTION = \ |
|
""" |
|
Wura is large-scale pretraining data for 20 languages popularly spoken in Africa. |
|
""" |
|
_HOMEPAGE = "https://github.com/castorini/AfriTeVa-keji" |
|
_LICENSE = "Apache License 2.0" |
|
_DOCUMENT_DATASET_VERSION = "1.0.0" |
|
_PASSAGE_DATASET_VERSION = "1.0.0" |
|
_LANGUAGES = { |
|
"Afrikaans": "afr", |
|
"Amharic": "amh", |
|
"Egyptian Arabic": "arz", |
|
"English": "eng", |
|
"French": "fra", |
|
"Hausa": "hau", |
|
"Igbo": "ibo", |
|
"Gahuza": "kin", |
|
"Malagasy": "mlg", |
|
"Chichewa": "nya", |
|
"Afaan Oromoo": "orm", |
|
|
|
"Portuguese": "por", |
|
"Shona": "sna", |
|
"Somali": "som", |
|
"Sesotho": "sot", |
|
"Swahili": "swa", |
|
"Tigrinya": "tir", |
|
"Xhosa": "xho", |
|
"Yoruba": "yor", |
|
"Zulu": "zul" |
|
} |
|
_DOCUMENT_DATASET_URL = Template("./documents-v1.0/${split}/${language}.jsonl") |
|
_PASSAGE_DATASET_URL = Template("./passages-v1.0/${split}/${language}.txt") |
|
|
|
INVALID_WINDOWS_CHARACTERS_IN_PATH = r"<>:/\|?*" |
|
|
|
|
|
@dataclass |
|
class WuraConfig(datasets.BuilderConfig): |
|
level: str = "document" |
|
|
|
|
|
class WuraDataset(datasets.GeneratorBasedBuilder): |
|
BUILDER_CONFIGS = [ |
|
WuraConfig( |
|
name=language, |
|
version=datasets.Version(_DOCUMENT_DATASET_VERSION), |
|
description=f"Wura dataset for language: {language}\n{_DESCRIPTION}", |
|
) for language in _LANGUAGES.values() |
|
] |
|
|
|
DEFAULT_CONFIG_NAME = "afr" |
|
|
|
def _info(self): |
|
if self.config.level == "document": |
|
features = ["id", "headline", "content", "category", "url"] |
|
elif self.config.level == "passage": |
|
features = ["id", "text"] |
|
else: |
|
raise ValueError("level can only be one of `document` or `passage`") |
|
|
|
features = {feature: datasets.Value("string") for feature in features} |
|
|
|
return datasets.DatasetInfo( |
|
description=self.config.description, |
|
features=datasets.Features(features), |
|
homepage=_HOMEPAGE, |
|
citation=_CITATION, |
|
license=_LICENSE |
|
) |
|
|
|
def _split_generators(self, dl_manager: DownloadManager): |
|
if self.config.level == "document": |
|
data_files = { |
|
split: _DOCUMENT_DATASET_URL.substitute( |
|
split=split, |
|
language=self.config.name, |
|
) for split in ["train", "eval"] |
|
} |
|
elif self.config.level == "passage": |
|
data_files = { |
|
split: _PASSAGE_DATASET_URL.substitute( |
|
split=split, |
|
language=self.config.name, |
|
) for split in ["train", "eval"] |
|
} |
|
else: |
|
raise ValueError("level can only be one of `document` or `passage`") |
|
|
|
language_files = dl_manager.download_and_extract(data_files) |
|
|
|
splits = [ |
|
datasets.SplitGenerator( |
|
name=datasets.Split.TRAIN, |
|
gen_kwargs={"filepath": language_files["train"]} |
|
), |
|
datasets.SplitGenerator( |
|
name=datasets.Split.VALIDATION, |
|
gen_kwargs={"filepath": language_files["eval"]} |
|
) |
|
] |
|
return splits |
|
|
|
def _generate_examples(self, filepath: str): |
|
with open(filepath, encoding="utf-8") as f: |
|
for idx, line in enumerate(f): |
|
if self.config.level == "document": |
|
data = json.loads(line) |
|
data["id"] = idx |
|
else: |
|
data = {"id": idx, "text": line.strip()} |
|
|
|
yield idx, data |
|
|