Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

korean-pretrain-training-data

한국어 · English · 简体中文 · 繁體中文 · 日本語 · Français · Deutsch · Español · Português (Brasil) · हिन्दी · العربية · עברית · Русский

한국어

개요

현재 3B 본학습 설정이 참조하는 실제 토큰 입력입니다. train 726파일과 고정 평가 6파일의 원래 바이트·순서·소스 가중치·seed·sampler v2를 보존했습니다. 이미 소비한 토큰만 추출한 로그나 사전에 섞은 단일 시퀀스가 아닙니다. 샘플링은 복원추출입니다.

문서 번역 언어는 학습 데이터에 그 언어가 포함된다는 뜻이 아닙니다. 주 자연언어는 한국어·영어이며 웹·수학·코드에 다른 언어가 섞일 수 있습니다.

구성과 규모

train 파일 726
고정 평가 파일 6
train tokens 121,174,609,758
고정 평가 tokens 14,124,617
token bytes 242,377,468,750

처리와 분리

평가 입력은 config의 명시적6개 holdout이 mix의726개 fallback 목록보다 우선합니다. holdout 목록을 지우지 마세요. 미포함720개는 전체 corpus에 있습니다. 형식은 little-endian uint16, 문서마다 EOS1회, EOS=1, vocab65,536, sampler=v2, seq_len4,096입니다. 토크나이저/16개 특수ID와 원mix·원config는 보존하며 다른 tokenizer는 원문을 재토큰화해야 합니다.

사용법

아래 명령은 저장소 루트에서 실행합니다. 새 출력/다운로드 디렉터리를 사용하고 다운로드 전에 범위·용량·고정 revision을 확인하세요. 합성 CPU 예제는 외부 데이터를 받지 않습니다. 실제 데이터는 별도 배포에서 필요한 파일만 선택하고 CHECKSUMS.json 또는 소스별 manifest의 크기·SHA256을 검증하세요.

python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
from huggingface_hub import snapshot_download
root = snapshot_download('pathcosmos/korean-pretrain-training-data', repo_type="dataset",
    revision='v1.0.1', allow_patterns=['data/tokens/final_v1/*', 'config/*', 'tokenizer/*', 'CHECKSUMS.json', 'TRAINING_GUIDE.md', 'release.json'],
    local_dir='korean-pretrain-training-data', token=False)
from pathlib import Path
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer
cache = snapshot_download("pathcosmos/korean-pretrain-training-data",
    repo_type="dataset", revision="v1.0.1",
    allow_patterns=["tokenizer/*"], token=False)
tokenizer = AutoTokenizer.from_pretrained(Path(cache) / "tokenizer")

TRAINING_GUIDE.md · config/main.original.json · data/tokens/final_v1/mix.json

모델과 학습

모델은 RMSNorm·RoPE·GQA·SwiGLU를 사용하는 순수 PyTorch Dense decoder입니다. 실제 3B 경로는 FSDP2·8GPU·eager이며 문맥4096, micro batch4, global batch2,097,152 tokens, peak LR3e-4, 목표57,000step입니다. 예제의 CPU micro 모델과 실제 분산 학습 경로는 다릅니다. 체크포인트 재개에는 모델·optimizer·RNG·입력/설정 계약 검증이 필요합니다.

소스 wiki_ko / fineweb2_ko / hplt_ko / fineweb_edu / owm / github_code_v1
소스 가중치 0.005 / 0.29464577095297395 / 0.20035422904702602 / 0.3 / 0.1 / 0.1

최적화·벤치마크

worker·thread·batch·SQLite 캐시·RAM·로컬/NFS와 eager/compile·정밀도·GPU 메모리를 비교했습니다. 총시간·처리량·peak 메모리·내용/수치 동등성을 함께 확인하고 채택·기각 이유를 기록합니다. 47개 캠페인·99개 결과 JSON을 제공하며 빠르더라도 수치 조건에 실패한 후보를 권장하지 않습니다. 짧은 표본의 최적값을 전체 최적값으로 주장하지 않습니다.

검증

GH 소스 공개판의 독립 CPU 검사는284통과/9건너뜀, 새 GitHub CPU 환경은258통과/35건너뜀입니다. 건너뜀은 원래 운영 기록 또는 별도 파일시스템이 필요한 검사입니다. 합성 quickstart는120문서·train100/eval20·vocab272·108,864parameter 모델로 저장/재개 후 model과optimizer 일치를 확인합니다. 데이터 배포의 내용 해시·분할/reader 검사와 전체3B 독립 재현은 구분합니다.

CHECKSUMS.json · release.json

출처와 라이선스

프로젝트 원본 코드는 MIT, 원본 설명문은 CC BY4.0입니다. 이는 원문 데이터의 권리를 대신하지 않습니다. 소스별 라이선스·귀속·고지·알려진 미확정 항목은 source.json, LICENSE-SCOPE.md와 출처 문서를 확인하세요. 모든 원문에 대한 무조건적 상업 재배포 허가를 주장하지 않습니다.

LICENSE-SCOPE.md · metadata/source_registry.json

한계

전역 유사중복 제거·개인정보/secret 완전 제거·공식 평가 오염 없음·최종 모델 품질을 보장하지 않습니다. 공개 시점의 본학습은 진행 중이며 SFT·모델 export·공식 언어능력 benchmark 완료를 뜻하지 않습니다. 과거 실험11개 코드 참조/고유9hash는 원본 snapshot을 회수하지 못했습니다. 일부 운영 도구는 별도 원래 실행 manifest가 필요합니다.

관련 저장소

GitHub · README GH / 사용법 · GH / 최적화·벤치마크 · GH / 검증

English

Overview

These are the actual token inputs referenced by the current 3B training configuration. The original bytes, order, source weights, seed and sampler v2 are preserved for726 training files and6 fixed evaluation files. This is neither a log of only consumed tokens nor a single pre-shuffled sequence. Sampling is with replacement.

Translation languages do not imply training-data coverage in those languages. The principal natural languages are Korean and English; web, math and code may contain other languages.

Contents and scale

Training files 726
Fixed evaluation files 6
Training tokens 121,174,609,758
Fixed evaluation tokens 14,124,617
Token bytes 242,377,468,750

Processing and splits

The6 explicit config holdouts take precedence over the726-file mix fallback. Do not clear the holdout list. The omitted720 files remain in the full corpus. Binaries are little-endian uint16 with one EOS per document, EOS=1, vocab65,536, sampler=v2 and seq_len4,096. Tokenizer bytes,16special IDs, original mix and config are preserved; a different tokenizer requires retokenizing the text.

Usage

Run commands below from the repository root. Use new output/download directories and check scope, size and pinned revision first. The synthetic CPU example downloads no external data. Select only the needed files from the separate data releases and verify sizes/SHA256 against CHECKSUMS.json or the source manifests.

python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
from huggingface_hub import snapshot_download
root = snapshot_download('pathcosmos/korean-pretrain-training-data', repo_type="dataset",
    revision='v1.0.1', allow_patterns=['data/tokens/final_v1/*', 'config/*', 'tokenizer/*', 'CHECKSUMS.json', 'TRAINING_GUIDE.md', 'release.json'],
    local_dir='korean-pretrain-training-data', token=False)
from pathlib import Path
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer
cache = snapshot_download("pathcosmos/korean-pretrain-training-data",
    repo_type="dataset", revision="v1.0.1",
    allow_patterns=["tokenizer/*"], token=False)
tokenizer = AutoTokenizer.from_pretrained(Path(cache) / "tokenizer")

TRAINING_GUIDE.md · config/main.original.json · data/tokens/final_v1/mix.json

Model and training

The model is a plain PyTorch dense decoder with RMSNorm, RoPE, GQA and SwiGLU. The actual3B path uses FSDP2,8GPUs and eager execution: context4096, micro batch4, global batch2,097,152tokens, peak LR3e-4 and target57,000steps. The CPU micro demonstration differs from distributed training. Exact resume requires model, optimizer, RNG and input/configuration contracts.

Source wiki_ko / fineweb2_ko / hplt_ko / fineweb_edu / owm / github_code_v1
Source weight 0.005 / 0.29464577095297395 / 0.20035422904702602 / 0.3 / 0.1 / 0.1

Optimization and benchmarks

Experiments compare workers, threads, batches, SQLite caches, RAM, local/NFS storage, eager/compile execution, precision and GPU memory. Total time, throughput, peak memory and content/numerical equivalence accompany adoption or rejection decisions.47 campaigns and99 result JSON files are included. Speed alone does not justify a numerically failing candidate, and a small sample does not establish a full-scale optimum.

Validation

The GH source release passed284 independent CPU tests with9 skips; fresh GitHub CPU CI passed258 with35 skips. Skips require original-run records or a second filesystem. The synthetic quickstart uses120documents, train100/eval20, vocab272 and a108,864parameter model, verifying exact model/optimizer resume. Dataset content-hash, split and reader checks are separate from full independent3B reproduction.

CHECKSUMS.json · release.json

Sources and licensing

Original project code is MIT and original prose is CC BY4.0. These grants do not replace original-content rights. Consult source.json, LICENSE-SCOPE.md and provenance documents for source-specific conditions, attribution, notices and known unresolved items. Unrestricted commercial redistribution of every original text is not asserted.

LICENSE-SCOPE.md · metadata/source_registry.json

Limitations

Global near-duplicate removal, complete PII/secret removal, contamination-free official evaluation and final model quality are not guaranteed. Main training was ongoing at the publication snapshot; SFT, model export and official language benchmarks are not completed features.11 historical code references/9unique hashes lack a recovered snapshot. Some operational tools require original-run manifests.

Related repositories

GitHub · README GH / Usage · GH / Optimization and benchmarks · GH / Validation

Downloads last month
226