BramVanroy/wikipedia_culturax_dutch

Filtered CulturaX + Wikipedia for Dutch

This is a combined and filtered version of CulturaX and Wikipedia, only including Dutch. It is intended for the training of LLMs.

Different configs are available based on the number of tokens (see a section below with an overview). This can be useful if you want to know exactly how many tokens you have. Great for using as a streaming dataset, too. Tokens are counted as white-space tokens, so depending on your tokenizer, you'll likely end up with more tokens than indicated here.

Every config also has a test set (for validation) of 1% the total size of the dataset, minimally 1 max. 64k samples (~16M tokens).

Wikipedia and CulturaX were shuffled before merging and the test set creation was also shuffled. Priority is given to Wikipedia to prioritize knowledge and cultural content, so the smaller configs will consist exclusively of Wikipedia and for the larger configs we augment with CulturaX. Every config builds further on the previous, so this means that every config contains the same data as the smaller ones and more HOWEVER their train/test splits are not the same, so test set of one config may overlap with samples for another training set. This is usually not a problem but just be aware that you do not train on one config's training set and test with another config's test set.

Citation

If you use Fietje or the CulturaX + Wikipedia filtered subset in your work, please cite to the following paper:

@misc{vanroy2024fietjeopenefficientllm,
      title={Fietje: An open, efficient LLM for Dutch}, 
      author={Bram Vanroy},
      year={2024},
      eprint={2412.15450},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.15450}, 
}

Configs

10k -- 79 samples -- 10,087 tokens

• ratio_wikipedia: 100.00%
• total_num_tokens: 10,087
• train_num_tokens: 9,205
• test_num_tokens: 882
• total_num_samples: 79
• train_num_samples: 78
• test_num_samples: 1

100k -- 1,057 samples -- 100,075 tokens

• ratio_wikipedia: 100.00%
• total_num_tokens: 100,075
• train_num_tokens: 98,044
• test_num_tokens: 2,031
• total_num_samples: 1,057
• train_num_samples: 1,047
• test_num_samples: 10

1M -- 10,802 samples -- 1,000,239 tokens

• ratio_wikipedia: 100.00%
• total_num_tokens: 1,000,239
• train_num_tokens: 991,119
• test_num_tokens: 9,120
• total_num_samples: 10,802
• train_num_samples: 10,694
• test_num_samples: 108

10M -- 141,263 samples -- 10,000,022 tokens

• ratio_wikipedia: 100.00%
• total_num_tokens: 10,000,022
• train_num_tokens: 9,874,772
• test_num_tokens: 125,250
• total_num_samples: 141,263
• train_num_samples: 139,851
• test_num_samples: 1,412

100M -- 1,028,484 samples -- 100,000,047 tokens

• ratio_wikipedia: 100.00%
• total_num_tokens: 100,000,047
• train_num_tokens: 99,013,372
• test_num_tokens: 986,675
• total_num_samples: 1,028,484
• train_num_samples: 1,018,200
• test_num_samples: 10,284

1B -- 5,153,898 samples -- 1,000,000,187 tokens

• ratio_wikipedia: 61.21%
• total_num_tokens: 1,000,000,187
• train_num_tokens: 989,990,190
• test_num_tokens: 10,009,997
• total_num_samples: 5,153,898
• train_num_samples: 5,102,360
• test_num_samples: 51,538

5B -- 20,833,009 samples -- 5,000,000,076 tokens

• ratio_wikipedia: 25.35%
• total_num_tokens: 5,000,000,076
• train_num_tokens: 4,984,493,654
• test_num_tokens: 15,506,422
• total_num_samples: 20,833,009
• train_num_samples: 20,769,009
• test_num_samples: 64,000

10B -- 40,240,566 samples -- 10,000,000,115 tokens

• ratio_wikipedia: 18.41%
• total_num_tokens: 10,000,000,115
• train_num_tokens: 9,984,156,828
• test_num_tokens: 15,843,287
• total_num_samples: 40,240,566
• train_num_samples: 40,176,566
• test_num_samples: 64,000

15B -- 59,648,123 samples -- 15,000,000,154 tokens

• ratio_wikipedia: 15.98%
• total_num_tokens: 15,000,000,154
• train_num_tokens: 14,983,970,518
• test_num_tokens: 16,029,636
• total_num_samples: 59,648,123
• train_num_samples: 59,584,123
• test_num_samples: 64,000

20B -- 79,055,679 samples -- 20,000,000,009 tokens

• ratio_wikipedia: 14.75%
• total_num_tokens: 20,000,000,009
• train_num_tokens: 19,983,799,357
• test_num_tokens: 16,200,652
• total_num_samples: 79,055,679
• train_num_samples: 78,991,679
• test_num_samples: 64,000

25B -- 98,463,236 samples -- 25,000,000,048 tokens

• ratio_wikipedia: 14.00%
• total_num_tokens: 25,000,000,048
• train_num_tokens: 24,983,765,326
• test_num_tokens: 16,234,722
• total_num_samples: 98,463,236
• train_num_samples: 98,399,236
• test_num_samples: 64,000

30B -- 117,870,793 samples -- 30,000,000,087 tokens

• ratio_wikipedia: 13.50%
• total_num_tokens: 30,000,000,087
• train_num_tokens: 29,983,707,932
• test_num_tokens: 16,292,155
• total_num_samples: 117,870,793
• train_num_samples: 117,806,793
• test_num_samples: 64,000

35B -- 137,278,350 samples -- 35,000,000,126 tokens

• ratio_wikipedia: 13.14%
• total_num_tokens: 35,000,000,126
• train_num_tokens: 34,983,914,739
• test_num_tokens: 16,085,387
• total_num_samples: 137,278,350
• train_num_samples: 137,214,350
• test_num_samples: 64,000

40B -- 156,685,907 samples -- 40,000,000,165 tokens

• ratio_wikipedia: 12.87%
• total_num_tokens: 40,000,000,165
• train_num_tokens: 39,983,508,625
• test_num_tokens: 16,491,540
• total_num_samples: 156,685,907
• train_num_samples: 156,621,907
• test_num_samples: 64,000

45B -- 176,093,463 samples -- 45,000,000,020 tokens

• ratio_wikipedia: 12.66%
• total_num_tokens: 45,000,000,020
• train_num_tokens: 44,983,608,118
• test_num_tokens: 16,391,902
• total_num_samples: 176,093,463
• train_num_samples: 176,029,463
• test_num_samples: 64,000

50B -- 195,501,020 samples -- 50,000,000,059 tokens

• ratio_wikipedia: 12.49%
• total_num_tokens: 50,000,000,059
• train_num_tokens: 49,983,567,461
• test_num_tokens: 16,432,598
• total_num_samples: 195,501,020
• train_num_samples: 195,437,020
• test_num_samples: 64,000

55B -- 214,908,577 samples -- 55,000,000,098 tokens

• ratio_wikipedia: 12.35%
• total_num_tokens: 55,000,000,098
• train_num_tokens: 54,983,723,278
• test_num_tokens: 16,276,820
• total_num_samples: 214,908,577
• train_num_samples: 214,844,577
• test_num_samples: 64,000

Filtering

While CultruaX already has done a lot of filtering, some more filtering can be done to improve the quality of the corpus. These filters are described below.

The baseline ratios (punctuation, uppercase, digits) were calculated on the SONAR-500 corpus (excluding WRPEA WRPED WRUEA WRUED WRUEB).

CulturaX:

• removed documents that contain the text "rechten voorbehouden" or "rights reserved"
• remove documents whose URL contained "wikipedia.org" (because we include a cleaned version of Wikipedia ourselves)
• removed documents that contain a "bad word" (see the section below)
• removed documents that contain any non-latin characters. The idea is that "knowledge"-based information (e.g. original writing of a name) are allowed when the data comes from Wikipedia, but not from any other webcrawl, to avoid unsollicited noise.

CulturaX + Wikipedia:

• removed documents where ratio of punctuation marks vs. non-whitespace characters is higher than 0.2
• removed documents where ratio of uppercase vs. non-whitespace characters is higher than 0.22
• removed documents where ratio of digits vs. non-whitespace characters is higher than 0.16
• removed documents where the average token length is < 2 or > 20

Bad words

BAD_PHRASES_DOC_LEVEL = {
    # https://en.wikipedia.org/wiki/Dutch_profanity
    "achterlijk",
    "debiel",
    "downie",
    "idioot",
    "kankerlijer",
    "klere",
    "kolere",
    "minkukel",
    "pestkop",
    "pleuris",
    "pleuritis",
    "teringlijer",
    "tyfuslijer",
    "gadver",
    "getver",
    "godver",
    "godskolere",
    "godverork",
    "graftak",
    "kopvod",
    "verdomme",
    "anaalgeneraal",
    "bitch",
    "dikzak",
    "flikker",
    "fok",
    "fuck",
    "hoer",
    "klootzak",
    "klote",
    "kreng",
    "kringspiermusketier",
    "kut",
    "lamzak",
    "lul",
    "manwijf",
    "matennaai",
    "neuken",
    "neuker",
    "ouwehoer",
    "reet",
    "reetkever",
    "reetridder",
    "rotzak",
    "schijt",
    "shit",
    "slet",
    "slijmbal",
    "slons",
    "sodemieter",
    "stoephoer",
    "swaffel",
    "teef",
    "trut",
    "tut",
    "zak",
    "uilskuiken",
    "zeik",
    "bamivreter",
    "bosneger",
    "neger",
    "fransoos",
    "geitenneuker",
    "kaaskop",
    "kakker",
    "koelie",
    "lijp",
    "medelander",
    "mocro",
    "mof",
    "nikker",
    "poepchinees",
    "roetmop",
    "spaghettivreter",
    "loempiavouwer",
    "spanjool",
    "spleetoog",
    "tatta",
    "tokkie",
    "zandneger",
    "zwartzak",
    "halvezool",
    "kenau",
    "klootviool",
    "knuppel",
    "koekert",
    "koekwaus",
    "oelewapper",
    "smeerlap",
    "sukkel",
    "sul",
    "wappie",
    "wijf",
    "zooi",
    # xxx (a.o. https://gitlab.com/yhavinga/c4nlpreproc/-/blob/master/clean/badwords_ennl.py?ref_type=heads)
    "xxx",
    "anal",
    "blowjob",
    "buttplug",
    "cock",
    "cunt",
    "geil",
    "sex",  # Standaardnederlands = seks, maybe we catch some porn or socialmedia sites with this misspelling
    "porn",
    # extra
    "nigger",
    "nigga",
    "hoerig",
    "klojo",
}

Config details

License information

For CulturaX: https://huggingface.co/datasets/uonlp/CulturaX#license-information For Wikipedia: https://huggingface.co/datasets/wikimedia/wikipedia#licensing-information