mkd-chanwoo/normalized-datasets-for-koreanLLM

Normalized Datasets for Korean LLM

Stage 0.5 — Normalization output of the Keural Korean LLM pretraining pipeline. Raw text from 19 source datasets converted into a single unified JSONL schema. No filtering has been applied. This is the clean, structured raw form.

---

Quick Stats

Metric	Value
Total documents normalized	~534,984,997
Number of source datasets	19
Domains	English · Korean · Code · Science
Format	JSONL (one JSON object per line)
Schema version	v1
Pipeline stage	Stage 0.5 (after download, before filtering)
Last updated	2026-04-09

---

Where This Fits in the Pipeline

flowchart LR
    A["Stage 0\nRaw Download\n~1.5 TB"] --> B["Stage 0.5\nNormalization\n← YOU ARE HERE\n~535M docs"]
    B --> C["Stage 1\nFiltering\nmkd-chanwoo/filtered-datasets-for-koreanLLM\n~293M docs"]
    C --> D["Stage 2\nDedup + Shard\nmkd-chanwoo/keural-datasets\n~329M docs / ~220B tokens"]
    style B fill:#f0a500,color:#000

---

What Is Normalization? (For Beginners)

Each source dataset stores text in a different format and field name. For example, Gutenberg uses the field name TEXT, but CCNews uses plain_text, and StarCoderData uses content.

Normalization solves this by:

1. Reading each source dataset in its original format
2. Extracting the text from its specific field
3. Writing everything into one unified schema with the same field names
4. Adding metadata like domain, language, doc_id, license

After normalization, all downstream stages (filtering, deduplication) work on one consistent format regardless of the original source.

---

Source Datasets & Field Mappings

Each row shows: what dataset was used, where it came from, which field was extracted as text, and how many documents were written after normalization.

English Domain

Dataset Key	HuggingFace Source	Source Field → text	Docs Normalized	Language
gutenberg	sedthh/gutenberg_english	TEXT	48,284	en
openwebtext	Skylion007/openwebtext	text	8,013,769	en
ccnews	stanford-oval/ccnews	plain_text	82,416,441	en
falcon-refinedweb	tiiuae/falcon-refinedweb	content	78,887,040	en
fineweb	HuggingFaceFW/fineweb	text	81,595,324	en
wikipedia	wikimedia/wikipedia config: 20231101.en	text	6,407,814	en

Korean Domain

Dataset Key	Source	Source Field → text	Docs Normalized	Language
namuwiki	heegyu/namuwiki-extracted	text	565,293	ko
wikipedia_ko	lcw99/wikipedia-korean-20240501	text	515,425	ko
oscar_ko_only	lcw99/oscar-ko-only	text	3,675,420	ko
korean_webtext	HAERAE-HUB/KOREAN-WEBTEXT	text	1,284,878	ko
aihub_modu	AIHub — Korean government open data (local)	parsed from structured AIHUB format	58,997	ko
aihub_books	AIHub — Korean government open data (local)	parsed from structured AIHUB format	5,823	ko
aihub_online_colloquial	AIHub — Korean government open data (local)	parsed from structured AIHUB format	22,859	ko

Code Domain

Dataset Key	HuggingFace Source	Source Field → text	Docs Normalized	Language
github-top-code	ronantakizawa/github-top-code	content	1,121,474	en (code)
codeparrot_clean	codeparrot/codeparrot-clean	content	5,361,374	en (code)
starcoderdata	bigcode/starcoderdata	content	104,640,054	en (code)

Science Domain

Dataset Key	HuggingFace Source	Source Field → text	Docs Normalized	Language
arxiv	KiteFishAI/arxiv-tex-corpus-full	text	1,089,469	en
open-web-math	open-web-math/open-web-math	text	6,315,233	en
peS2o	allenai/peS2o	text	151,960,046	en

---

Total Documents by Domain

Domain	Docs Normalized	% of Total
English	~257,368,672	48.1%
Science	~159,364,748	29.8%
Code	~111,122,902	20.8%
Korean	~6,128,675	1.1%
Total	~534,984,997	100%

---

Normalization Process — Step by Step

flowchart TD
    A["Source File\n(HuggingFace / AIHub)"] --> B["Read original format\n(JSONL / Parquet / CSV / AIHUB)"]
    B --> C["Extract text from\ndataset-specific field\n(TEXT / text / content / plain_text)"]
    C --> D["Assign doc_id\n= source_name + '_' + index"]
    D --> E["Compute char_count\nand tokens_count\n(keural SentencePiece tokenizer)"]
    E --> F["Write unified JSONL\nwith full metadata"]
    F --> G["Update checkpoint\n(resumable mid-stream)"]

Normalization is resumable. Each dataset tracks line_index and doc_id in a checkpoint file (stage1_savepoint.json), so if the process is interrupted, it resumes exactly where it left off.

What normalization does NOT do:

• ❌ Does not modify or clean text content
• ❌ Does not filter documents
• ❌ Does not deduplicate
• ❌ Does not re-encode or translate
• ✅ Only reshapes structure and adds metadata

---

Unified Document Schema

Every document in this repository follows this exact 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 (if available, else null)",
  "license":            "Public Domain",
  "source_file":        "data/raw/gutenberg/train-00000-of-00001.parquet",
  "source_index":       42,
  "timestamp":          "2026-03-15T08:22:11Z  (if available in source, else null)",
  "processing_version": "v1"
}

Field Descriptions

Field	Type	Description
doc_id	string	Unique ID: {source_name}_{source_index}
source_name	string	Dataset key (e.g. gutenberg, ccnews)
domain	string	One of: english, korean, code, science
language	string	ISO 639-1 code: en or ko
text	string	Raw document text (unmodified from source)
url	string|null	Original URL if provided by source dataset
license	string	Source dataset license
source_file	string	Local path to the source file it was read from
source_index	int	Row index within that source file
timestamp	string|null	Publication date/time if available in source
processing_version	string	Pipeline version (v1)

---

Normalization Statistics (Seen vs Written)

"Seen" = total rows read from source. "Written" = successfully normalized. Rows not written are rows that failed to parse (malformed JSON, empty text, encoding errors).

Dataset	Seen	Written	Normalized Rate
aihub_books	5,974	5,823	97.5%
aihub_modu	117,994	58,997	50.0% (deduped at read)
aihub_online_colloquial	45,894	22,859	49.8% (deduped at read)
arxiv	1,089,469	1,089,469	100%
ccnews	82,416,441	82,416,441	100%
codeparrot_clean	5,361,374	5,361,374	100%
falcon-refinedweb	78,888,470	78,887,040	~100%
fineweb	81,595,324	81,595,324	100%
github-top-code	1,122,139	1,121,474	~100%
gutenberg	48,285	48,284	~100%
korean_webtext	1,284,879	1,284,878	~100%
namuwiki	565,293	565,293	100%
open-web-math	6,315,233	6,315,233	100%
openwebtext	8,013,769	8,013,769	100%
oscar_ko_only	3,675,421	3,675,420	~100%
peS2o	151,960,046	151,960,046	100%
starcoderdata	104,640,054	104,640,054	100%
wikipedia (en)	6,407,814	6,407,814	100%
wikipedia_ko	515,425	515,425	100%

---

Tokenizer Used for Token Counting

All tokens_count values in this dataset are computed using the Keural SentencePiece tokenizer:

• Model: mkd-ai/keural-tokenizer
• Type: SentencePiece (Unigram)
• Vocabulary file: keural_tokenizer.vocab
• Model file: keural_tokenizer.model

This is the same tokenizer used by the Keural LLM model.

---

Download & Processing Timeline

Event	Date (KST)
Download of first datasets begins	2026-04-01
Normalization of first batch complete	2026-04-08
All 19 datasets normalized	2026-04-09
Upload to this HuggingFace repo	2026-04-09 ~09:43 KST
Last updated	2026-04-10

---

Raw Download Sizes (Stage 0)

Dataset	Raw Download Size
peS2o	286.58 GB
aihub_specialized_corpus	166.18 GB
ccnews	148.52 GB
starcoderdata	149.65 GB
fineweb	149.97 GB
falcon-refinedweb	127.93 GB
open-web-math	25.55 GB
openwebtext	22.53 GB
aihub_books	110.98 GB
aihub_modu	38.16 GB
aihub_online_colloquial	17.66 GB
codeparrot_clean	11.93 GB
wikipedia (en)	10.83 GB
gutenberg	10.01 GB
oscar_ko_only	6.49 GB
korean_webtext	4.17 GB
wikipedia_ko	1.63 GB

---

Licenses

This dataset contains content from multiple sources with mixed licenses. Each source retains its original license.

Dataset	License
gutenberg	Public Domain
openwebtext	CC0 1.0
ccnews	CC-BY 4.0
falcon-refinedweb	Falcon License (TII)
fineweb	ODC-By 1.0
wikipedia (en)	CC-BY-SA 3.0
namuwiki	CC-BY-NC-SA 3.0
wikipedia_ko	CC-BY-SA 3.0
oscar_ko_only	CC0 1.0
korean_webtext	CC-BY 4.0
aihub_modu	AIHub Open License (Korean government open data)
aihub_books	AIHub Open License
aihub_online_colloquial	AIHub Open License
github-top-code	Various open source (see source repo)
codeparrot_clean	OpenRAIL
starcoderdata	BigCode OpenRAIL-M
arxiv	CC-BY 4.0
open-web-math	CC-BY 4.0
peS2o	CC-BY 4.0

⚠️ License Notice: This repository inherits mixed licenses from its source datasets. Please review the license of each individual source before commercial or research use.

---

Related Repositories

Repo	Stage	Description
This repo	Stage 0.5	Normalized raw data
mkd-chanwoo/filtered-datasets-for-koreanLLM	Stage 1	Quality + language + toxicity filtered
mkd-chanwoo/keural-datasets	Stage 2	Final deduplicated + sharded production data
mkd-chanwoo/simplemodel-270M	Model	LLM trained on this pipeline's output