README.md

metadata

license: cc-by-sa-4.0
task_categories:
  - text-generation
language:
  - sr
  - hr
  - bs
tags:
  - webdataset
pretty_name: Kišobran (Umbrella corp.)
size_categories:
  - 10B<n<100B
configs:
  - config_name: default
    data_files:
      - split: train
        path: '*.txt'
      - split: sr
        path: '*_sr.txt'
      - split: cnr
        path: '*_cnr.txt'
      - split: hr
        path: '*_hr.txt'
      - split: bs
        path: '*_bs.txt'

Kišobran korpus - krovni veb korpus srpskog i srpskohrvatskog jezika Najveća agregacija veb korpusa do sada, pogodna za obučavanje velikih jezičkih modela za srpski jezik. Ukupno x dokumenata, ukupno sa preko 18.5 milijardi reči. Svaka linija predstavlja novi dokument Rečenice unutar dokumenata su obeležene. Sadrži obrađene i deduplikovane verzije sledećih korpusa:

Umbrella corp. - umbrella web corpus of Serbian and Serbo-Croatian The largest aggregation of web corpora so far, suitable for training Serbian large language models. A total of x documents containing over 18.5 billion words. Each line represents a document. Each Sentence in a document is delimited. Contains processed and deduplicated versions of the following corpora:

Korpus Corpus	Jezik Language	Broj dokumenata Doc. count	Broj reči Word count	Udeo Share
HPLT_sr	🇷🇸	2.9 M	2.5 B	13.74%
MaCoCu_sr	🇷🇸	6.7 M	2.1 B	11.54%
MC4_sr	🇷🇸	2.3 M	782 M	4.19%
cc100_sr	🇷🇸	2.3 M	659 M	3.53%
PDRS1.0	🇷🇸	400 K	506 M	2.71%
SrpKorNews	🇷🇸	35 K	469 M	2.51%
OSCAR_sr	🇷🇸	500 K	410 M	2.2%
srWaC	🇷🇸	1.2 M	307 M	1.65%
CLASSLA_sr	🇷🇸	1.3 M	240 M	1.29%
MaCoCu_cnr	🇷🇸/🇲🇪	500 K	152 M	0.82%
meWaC	🇷🇸/🇲🇪	200 K	41 M	0.22%
cc100_hr	🇭🇷	13.3 M	2.5 B	13.73%
MaCoCu_hr	🇭🇷	8 M	2.3 B	12.63%
HPLT_hr	🇭🇷	2.3 M	1.8 B	9.95%
hr_news	🇭🇷	4.1 M	1.4 B	7.65%
hrWaC	🇭🇷	3.1 M	935 M	5.01%
CLASSLA_hr	🇭🇷	1.2 M	160 M	0.86%
riznica	🇭🇷	20 K	69 M	0.37%
MaCoCu_bs	🇧🇦	2.6 M	700 M	3.75%
bsWaC	🇧🇦	800 K	194 M	1.04%
CLASSLA_bs	🇧🇦	800 K	105 M	0.56%
cc100_bs	🇧🇦	300 K	9 M	0.05%
TOTAL		54.75 M	18.65 B	100%

Load complete dataset / Učitavanje kopletnog dataseta

from datasets import load_dataset
dataset = load_dataset("procesaur/umbrella")

Load a specific language / Učitavanje pojedinačnih jezika

from datasets import load_dataset
dataset_sr = load_dataset("procesaur/umbrella", split="sr")
dataset_cnr = load_dataset("procesaur/umbrella", split="cnr")
dataset_hr = load_dataset("procesaur/umbrella", split="hr")
dataset_bs = load_dataset("procesaur/umbrella", split="bs")

Editor

Mihailo Škorić

@procesaur

Citation:

@article{skoric24korpusi,
  author    = {\vSkori\'c, Mihailo and Jankovi\'c, Nikola},
  title     = {New Textual Corpora for Serbian Language Modeling},
  journal   = {Infotheca},
  volume    = {24},
  issue     = {1},
  year      = {2024},
  publisher = {Zajednica biblioteka univerziteta u Srbiji, Beograd},
  url       = {https://arxiv.org/abs/2405.09250}
}

Istraživanje je sprovedeno uz podršku Fonda za nauku Republike Srbije, #7276, Text Embeddings – Serbian Language Applications – TESLA. Svaki korpus u tabeli vezan je za URL sa kojeg je preuzet. Prikazani brojevi dokumenata i reči, odnose se na stanje nakon čićenja i deduplikacije. Deduplikacija je izvršena pomoću alata onion korišćenjem pretrage 6-torki i pragom dedumplikacije 75%. Računarske resursre neophodne za deduplikaciju korpusa obezbedila je Nacionalna platforma za veštačku inteligenciju Srbije.

This research was supported by the Science Fund of the Republic of Serbia, #7276, Text Embeddings - Serbian Language Applications - TESLA. Each corpus in the table is linked to the URL from which it was downloaded. The displayed numbers of documents and words refer to after cleaning and deduplication. The dataset was deduplicated using onion using 6-tuples search and a duplicate threshold of 75%. Computer resources necessary for the deduplication of the corpus were provided by the National Platform for Artificial Intelligence of Serbia.