glue.py

import os
import json
import datasets
from datasets import BuilderConfig, Features, ClassLabel, Value, Sequence


_DESCRIPTION = """
# 한국어 지시학습 데이터셋
- glue 데이터셋을 한국어로 변역한 데이터셋
"""

_CITATION = """
@inproceedings{KITD,
  title={언어 번역 모델을 통한 한국어 지시 학습 데이터 세트 구축},
  author={임영서, 추현창, 김산, 장진예, 정민영, 신사임},
  booktitle={제 35회 한글 및 한국어 정보처리 학술대회},
  pages={591--595},
  month=oct,
  year={2023}
}
"""

# glue
_COLA_FEATURES = Features({
    "data_index_by_user": Value(dtype="int32"),
    "label": Value(dtype="int32"),
    "sentence": Value(dtype="string"),    
})

def _parsing_cola(file_path):
    with open(file_path, mode="r") as f:
        dataset = json.load(f)
    for _idx, data in enumerate(dataset):
        _data_index_by_user = data["data_index_by_user"]
        _label = data["label"]
        _sentence = data["sentence"]

        yield _idx, {
            "data_index_by_user": _data_index_by_user,
            "label": _label,
            "sentence": _sentence
        }

_MRPC_FEATURES = Features({
    "data_index_by_user": Value(dtype="int32"),
    "sentence1": Value(dtype="string"),
    "sentence2": Value(dtype="string"),
    "label": Value(dtype="int32"),
    "idx": Value(dtype="int32")
})

def _parsing_mrpc(file_path):
    with open(file_path, mode="r") as f:
        dataset = json.load(f)
    for _i, data in enumerate(dataset):
        _data_index_by_user = data["data_index_by_user"]
        _sentence1 = data["sentence1"]
        _sentence2 = data["sentence2"]
        _label = data["label"]
        _idx = data["idx"]

        yield _i, {
            "data_index_by_user": _data_index_by_user,
            "sentence1": _sentence1,
            "sentence2": _sentence2,
            "label": _label,
            "idx": _idx,
        }

_QNLI_FEATURES = Features({
    "data_index_by_user": Value(dtype="int32"),
    "label": Value(dtype="int32"),
    "question": Value(dtype="string"),
    "sentence": Value(dtype="string"),
})

def _parsing_qnli(file_path):
    with open(file_path, mode="r") as f:
        dataset = json.load(f)
    for _idx, data in enumerate(dataset):
        _data_index_by_user = data["data_index_by_user"]
        _label = data["label"]
        _question = data["question"]
        _sentence = data["sentence"]

        yield _idx, {
            "data_index_by_user": _data_index_by_user,
            "label": _label,
            "question": _question,
            "sentence": _sentence,
        }

_QQP_FEATURES = Features({
    "data_index_by_user": Value(dtype="int32"),
    "question1": Value(dtype="string"),
    "question2": Value(dtype="string"),
    "label": Value(dtype="int32"),
    "idx": Value(dtype="int32")
})

def _parsing_qqp(file_path):
    with open(file_path, mode="r") as f:
        dataset = json.load(f)
    for _i, data in enumerate(dataset):
        _data_index_by_user = data["data_index_by_user"]
        _question1 = data["question1"]
        _question2 = data["question2"]
        _label = data["label"]
        _idx = data["idx"]

        yield _i, {
            "data_index_by_user": _data_index_by_user,
            "question1": _question1,
            "question2": _question2,
            "label": _label,
            "idx": _idx,
        }

_WNLI_FEATURES = Features({
    "data_index_by_user": Value(dtype="int32"),
    "sentence1": Value(dtype="string"),
    "sentence2": Value(dtype="string"),
    "label": Value(dtype="int32"),
    "idx": Value(dtype="int32")
})

def _parsing_wnli(file_path):
    with open(file_path, mode="r") as f:
        dataset = json.load(f)
    for _i, data in enumerate(dataset):
        _data_index_by_user = data["data_index_by_user"]
        _sentence1 = data["sentence1"]
        _sentence2 = data["sentence2"]
        _label = data["label"]
        _idx = data["idx"]

        yield _i, {
            "data_index_by_user": _data_index_by_user,
            "sentence1": _sentence1,
            "sentence2": _sentence2,
            "label": _label,
            "idx": _idx,
        }

class GlueConfig(BuilderConfig):
    def __init__(self, name, feature, reading_fn, parsing_fn, citation, **kwargs):
        super(GlueConfig, self).__init__(
            name = name,
            version=datasets.Version("1.0.0"),
            **kwargs)
        self.feature = feature
        self.reading_fn = reading_fn
        self.parsing_fn = parsing_fn
        self.citation = citation

class GLUE(datasets.GeneratorBasedBuilder):
    BUILDER_CONFIGS = [
        GlueConfig(
            name = "cola",
            data_dir = "./glue",
            feature = _COLA_FEATURES,
            reading_fn = _parsing_cola,
            parsing_fn = lambda x:x,
            citation = _CITATION,
        ),
        GlueConfig(
            name = "mrpc",
            data_dir = "./glue",
            feature = _MRPC_FEATURES,
            reading_fn = _parsing_mrpc,
            parsing_fn = lambda x:x,
            citation = _CITATION,
        ),
        GlueConfig(
            name = "qnli",
            data_dir = "./glue",
            feature = _QNLI_FEATURES,
            reading_fn = _parsing_qnli,
            parsing_fn = lambda x:x,
            citation = _CITATION,
        ),
        GlueConfig(
            name = "qqp",
            data_dir = "./glue",
            feature = _QQP_FEATURES,
            reading_fn = _parsing_qqp,
            parsing_fn = lambda x:x,
            citation = _CITATION,
        ),
        GlueConfig(
            name = "wnli",
            data_dir = "./glue",
            feature = _WNLI_FEATURES,
            reading_fn = _parsing_wnli,
            parsing_fn = lambda x:x,
            citation = _CITATION,
        ),
    ]
    
    def _info(self) -> datasets.DatasetInfo:
        """Returns the dataset metadata."""
        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=self.config.feature,
            citation=_CITATION,
        )
    
    def _split_generators(self, dl_manager: datasets.DownloadManager):
        """Returns SplitGenerators"""
        if self.config.name == "qqp":
            path_kv = {
                datasets.Split.TRAIN:[
                    os.path.join(dl_manager.manual_dir, f"{self.config.name}/train.json")
                ],
            }
        else:
            path_kv = {
                datasets.Split.TRAIN:[
                    os.path.join(dl_manager.manual_dir, f"{self.config.name}/train.json")
                ],
                datasets.Split.VALIDATION:[
                    os.path.join(dl_manager.manual_dir, f"{self.config.name}/validation.json")
                ],
                datasets.Split.TEST:[
                    os.path.join(dl_manager.manual_dir, f"{self.config.name}/test.json")
                ],
            }
        return [
            datasets.SplitGenerator(name=k, gen_kwargs={"path_list": v})
            for k, v in path_kv.items()
        ]
    
    def _generate_examples(self, path_list):
        """Yields examples."""
        for path in path_list:
            try:
                for example in iter(self.config.reading_fn(path)):
                    yield self.config.parsing_fn(example)
            except Exception as e:
                print(e)