metadata

annotations_creators:
  - no-annotation
language_creators:
  - crowdsourced
language:
  - ko
license:
  - cc-by-4.0
multilinguality:
  - monolingual
pretty_name: laion2B-multi-korean-subset
size_categories:
  - 10M<n<100M
task_categories:
  - feature-extraction

laion2B-multi-korean-subset

Dataset Description

Homepage: laion-5b
Huggingface: laion/laion2B-multi

About dataset

Data organized by extracting only Korean data from laion/laion2B-multi

Lisence

CC-BY-4.0

Data Structure

Data Instance

>>> from datasets import load_dataset
>>> dataset = load_dataset("Bingsu/laion2B-multi-korean-subset")
>>> dataset
DatasetDict({
    train: Dataset({
        features: ['SAMPLE_ID', 'URL', 'TEXT', 'HEIGHT', 'WIDTH', 'LICENSE', 'LANGUAGE', 'NSFW', 'similarity'],
        num_rows: 11376263
    })
})

>>> dataset["train"].features
{'SAMPLE_ID': Value(dtype='int64', id=None),
 'URL': Value(dtype='string', id=None),
 'TEXT': Value(dtype='string', id=None),
 'HEIGHT': Value(dtype='int32', id=None),
 'WIDTH': Value(dtype='int32', id=None),
 'LICENSE': Value(dtype='string', id=None),
 'LANGUAGE': Value(dtype='string', id=None),
 'NSFW': Value(dtype='string', id=None),
 'similarity': Value(dtype='float32', id=None)}

Data Size

download: 1.56 GiB
generated: 2.37 GiB
total: 3.93 GiB

Data Field

'SAMPLE_ID': int
'URL': string
'TEXT': string
'HEIGHT': int
'WIDTH': int
'LICENSE': string
'LANGUAGE': string
'NSFW': string
'similarity': float

Data Splits

	train
# of texts	11376263

Note

Height, Width

모든 데이터를 살펴본 것은 아니지만, 이미지의 가로가 HEIGHT로, 세로가 WIDTH로 되어있는 것 같습니다.

>>> dataset["train"][98]
{'SAMPLE_ID': 2937471001780,
 'URL': 'https://image.ajunews.com/content/image/2019/04/12/20190412175643597949.png',
 'TEXT': '인천시교육청, 인천 시군구발전협의회  임원진과의 간담회 개최',
 'HEIGHT': 640,
 'WIDTH': 321,
 'LICENSE': '?',
 'LANGUAGE': 'ko',
 'NSFW': 'UNLIKELY',
 'similarity': 0.33347243070602417}

image

Code used to generate

import csv
import re

from datasets import load_dataset
from tqdm import tqdm


pattern = re.compile(r"[가-힣]")


def quote(s: str) -> str:
    s = s.replace('"""', "")
    return s


def filter_func(example) -> bool:
    lang = example.get("LANGUAGE")
    text = example.get("TEXT")
    if not isinstance(lang, str) or not isinstance(text, str):
        return False
    return lang == "ko" or pattern.search(text) is not None


file = open("./laion2B-mulit_korean_subset.csv", "w", encoding="utf-8", newline="")

ds = load_dataset("laion/laion2B-multi", split="train", streaming=True)
dsf = ds.filter(filter_func)
header = [
    "SAMPLE_ID",
    "URL",
    "TEXT",
    "HEIGHT",
    "WIDTH",
    "LICENSE",
    "LANGUAGE",
    "NSFW",
    "similarity",
]
writer = csv.DictWriter(file, fieldnames=header, delimiter="\t")
writer.writeheader()

try:
    for data in tqdm(dsf):
        data["TEXT"] = quote(data.get("TEXT", ""))
        if data["TEXT"]:
            writer.writerow(data)
finally:
    file.close()

print("Done!")

이후에 HEIGHT나 WIDTH가 None인 데이터를 제거하고 업로드하였습니다.

img2dataset

img2dataset을 사용하여 URL로된 이미지들을 데이터셋형태로 만들 수 있습니다.