mkd-chanwoo/normalized-datasets-for-koreanLLM

Normalized Datasets for Korean LLM (Stage 0.5)

This is a comprehensive pretraining dataset for Korean LLMs containing ~944M normalized documents from 42 source datasets across English, Korean, Code, and Science domains.

Quick Stats

Metric	Value
Total Documents	~944,545,628
Source Datasets	42
Format	JSONL (one JSON object per line)
Languages	English, Korean
Domains	English, Korean, Code, Science
Pipeline Stage	Stage 0.5 (after download, before filtering)
Last Updated	2026-06-02

Total Documents by Domain

Domain	Documents	% of Total
English	360,821,206	38.2%
Korean	254,422,153	26.9%
Code	136,706,183	14.5%
Science	192,596,086	20.4%
Total	944,545,628	100%

What is Normalization?

This dataset standardizes 42 source datasets with different formats into a unified JSONL schema. Each source uses different field names (e.g., TEXT, plain_text, content), so normalization:

1. Reads each source in its original format
2. Extracts text from its specific field
3. Writes to unified schema with consistent field names
4. Adds metadata (domain, language, doc_id, license)

Does NOT: modify text, filter, deduplicate, re-encode, or translate

Unified Document Schema

{
  "doc_id":             "gutenberg_000000042",
  "source_name":        "gutenberg",
  "domain":             "english",
  "language":           "en",
  "text":               "The full original text of the document...",
  "url":                "https://source-url.com/page",
  "license":            "Public Domain",
  "source_file":        "data/raw/gutenberg/train-00000-of-00001.parquet",
  "source_index":       42,
  "timestamp":          "2026-03-15T08:22:11Z",
  "processing_version": "v2"
}

Source Datasets & Field Mappings

Dataset	Domain	HuggingFace Repo	Text Field
gutenberg	english	sedthh/gutenberg_english	TEXT
ccnews	english	stanford-oval/ccnews	plain_text
falcon-refinedweb	english	tiiuae/falcon-refinedweb	content
fineweb	english	HuggingFaceFW/fineweb	text
fineweb_edu	english	HuggingFaceFW/fineweb-edu	text
hplt_english_1/2/3	english	HPLT/HPLT2.0_cleaned	text
openwebtext	english	Skylion007/openwebtext	text
wikipedia	english	wikimedia/wikipedia	text
namuwiki	korean	heegyu/namuwiki-extracted	text
wikipedia_ko	korean	lcw99/wikipedia-korean-20240501	text
oscar_ko_only	korean	lcw99/oscar-ko-only	text
korean_webtext	korean	HAERAE-HUB/KOREAN-WEBTEXT	text
hplt_korean	korean	HPLT/HPLT2.0_cleaned	text
culturax_ko	korean	uonlp/CulturaX	text
wanjuan_korean	korean	opendatalab/WanJuan-Korean	content
fineweb2_korean	korean	HuggingFaceFW/fineweb-2	text
c4_korean	korean	allenai/c4	text
cc100_documents_korean	korean	singletongue/cc100-documents	text
cc_korean	korean	chaannwooff/CC-korean	text
korean_web	korean	chaannwooff/korean-web	text
korean_law	korean	chaannwooff/Korean_law	text
dartdoc	korean	chaannwooff/Dartdoc	text
aihub_modu	korean	AIHub (aihubshell)	JSONPath extraction
aihub_books	korean	AIHub (aihubshell)	JSONPath extraction
aihub_online_colloquial	korean	AIHub (aihubshell)	JSONPath extraction
aihub_korean_corpus_literature	korean	AIHub (aihubshell)	JSONPath extraction
github-top-code	code	ronantakizawa/github-top-code	content
starcoderdata	code	bigcode/starcoderdata	content
codeparrot_clean	code	codeparrot/codeparrot-clean	content
the_stack_c	code	bigcode/the-stack	content
the_stack_java	code	bigcode/the-stack	content
the_stack_python	code	bigcode/the-stack	content
arxiv	science	KiteFishAI/arxiv-tex-corpus-full	text
open-web-math	science	open-web-math/open-web-math	text
peS2o	science	allenai/peS2o	text
algebraic_stack	science	typeof/algebraic-stack	text
fiwebmath_4plus	science	HuggingFaceTB/finemath	text
open_med_text	science	ywchoi/OpenMedText	text
pubmed_abstracts	science	casinca/PUBMED_title_abstracts_2019_baseline	text
s2orc	science	sentence-transformers/s2orc	abstract

Normalization Statistics (Seen vs Written)

Datasets processed by the Stage 0.5 normalizer (tracked in checkpoint). Datasets not listed below were normalized via legacy path.

Dataset	Seen	Written	Rate
algebraic_stack	3,440,694	3,440,226	100.0%
c4_korean	15,618,718	15,618,718	100.0%
cc100_documents_korean	35,678,358	35,678,358	100.0%
cc_korean	3,570,675	3,570,675	100.0%
culturax_ko	20,557,310	20,557,309	100.0%
dartdoc	256,548	256,548	100.0%
fineweb	81,595,324	81,595,324	100.0%
fineweb2_korean	60,904,429	60,904,429	100.0%
fineweb_edu	57,121,167	57,121,167	100.0%
fiwebmath_4plus	6,699,493	6,699,493	100.0%
hplt_english_1	37,128,244	37,128,244	100.0%
hplt_english_2	16,085,779	16,085,779	100.0%
hplt_english_3	22,951,515	22,951,515	100.0%
hplt_korean	38,866,835	38,866,835	100.0%
korean_law	2,399,006	2,399,006	100.0%
korean_web	1,545,735	1,545,735	100.0%
korean_webtext	1,284,879	1,284,878	100.0%
open_med_text	127,736	127,736	100.0%
oscar_ko_only	3,675,421	3,675,420	100.0%
pubmed_abstracts	15,518,009	15,517,555	100.0%
s2orc	89,455,939	89,455,939	100.0%
the_stack_c	21,383,832	21,383,810	100.0%
the_stack_java	42,429,211	42,429,204	100.0%
the_stack_python	24,214,270	24,214,204	100.0%
wanjuan_korean	68,894,938	68,894,937	100.0%
wikipedia	6,407,814	6,407,814	100.0%
aihub_korean_corpus_literature	3,276,057	908	0.0% ⚠️

⚠️ aihub_korean_corpus_literature: raw JSON 구조 미스매치로 정규화율 매우 낮음. 추후 extraction path 수정 필요.

Key Features

• Resumable Processing: Checkpoint system allows resuming interrupted normalization
• High Normalization Rate: Most datasets achieve >99% normalization (minimal parsing failures)
• Mixed Licenses: Includes public domain, CC, MIT, Apache, and restricted licenses
• Same Tokenizer: Uses Keural SentencePiece tokenizer (mkd-ai/keural-tokenizer) for token counting

Pipeline Context

Stage 0: Raw Download (~1.5 TB)
    ↓
Stage 0.5: Normalization → ~944M docs (YOU ARE HERE)
    ↓
Stage 1: Filtering → ~704M docs
    ↓
Stage 2: Dedup + Shard → ~512B tokens

License Considerations

Commercial use varies by source:

• ✅ OK: FineWeb, Wikipedia, OpenWebText, C4, arXiv, PubMed (with conditions)
• ⚠️ Caution: CCNews, Falcon RefinedWeb, some medical sources
• ❌ No Commercial: NamuWiki (CC BY-NC-SA), AIHub datasets (research only)
• 🔒 Restricted: StarCoderData, The Stack (BigCode OpenRAIL-M)

Review individual source licenses before commercial use.