Datasets:
Document13languages with complete bilingual root and reciprocal GitHub navigation
Browse files- DOCUMENTATION_CHECKSUMS.json +80 -0
- README.md +199 -46
- cards/README.ar.md +111 -0
- cards/README.de.md +111 -0
- cards/README.en.md +111 -0
- cards/README.es.md +111 -0
- cards/README.fr.md +111 -0
- cards/README.he.md +111 -0
- cards/README.hi.md +111 -0
- cards/README.ja.md +111 -0
- cards/README.ko.md +111 -0
- cards/README.pt-BR.md +111 -0
- cards/README.ru.md +111 -0
- cards/README.zh-CN.md +111 -0
- cards/README.zh-TW.md +111 -0
- documentation.json +46 -0
DOCUMENTATION_CHECKSUMS.json
ADDED
|
@@ -0,0 +1,80 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"scope": "Documentation only; data hashes remain in manifest.json",
|
| 3 |
+
"files": [
|
| 4 |
+
{
|
| 5 |
+
"path": "README.md",
|
| 6 |
+
"bytes": 17916,
|
| 7 |
+
"sha256": "cfe3bde40fff25ce20b5dea83500386a56e6090faaa0a25db68faaa2d932465a"
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"path": "cards/README.ar.md",
|
| 11 |
+
"bytes": 11972,
|
| 12 |
+
"sha256": "6ab594363cd764c5a67a9756562f8d215eaa648a1c5dadbd7a431193ba90045a"
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"path": "cards/README.de.md",
|
| 16 |
+
"bytes": 10772,
|
| 17 |
+
"sha256": "f3d001af1f5b6941d44782edbdaa1ec6a419f2bd47f5d793ca288fafad6d219d"
|
| 18 |
+
},
|
| 19 |
+
{
|
| 20 |
+
"path": "cards/README.en.md",
|
| 21 |
+
"bytes": 10699,
|
| 22 |
+
"sha256": "d9119539f72eaa629d6c978b3870908ec19644e562e06f0cedda44c8c8b832e1"
|
| 23 |
+
},
|
| 24 |
+
{
|
| 25 |
+
"path": "cards/README.es.md",
|
| 26 |
+
"bytes": 10528,
|
| 27 |
+
"sha256": "b1f9e404bd4b89ecea8b1fe02aeb2f51560398c7959fbe863459120657eff5c1"
|
| 28 |
+
},
|
| 29 |
+
{
|
| 30 |
+
"path": "cards/README.fr.md",
|
| 31 |
+
"bytes": 10880,
|
| 32 |
+
"sha256": "26506da92451d720036617d63ae273f00b37da1f8042ada00a1eac6ea4c6ce4f"
|
| 33 |
+
},
|
| 34 |
+
{
|
| 35 |
+
"path": "cards/README.he.md",
|
| 36 |
+
"bytes": 11354,
|
| 37 |
+
"sha256": "1e6766605325638fc3ef839ae26cb950064bc7dacfa5b35802f56c32338f4560"
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"path": "cards/README.hi.md",
|
| 41 |
+
"bytes": 13756,
|
| 42 |
+
"sha256": "bcc5ebf71fef16dad1921f42bb5c90749ae36fee795f0a15315a5779f3a24e94"
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"path": "cards/README.ja.md",
|
| 46 |
+
"bytes": 10740,
|
| 47 |
+
"sha256": "717ee077f6d1817365172df4f5d97ea84b978a96a32bd7618f60c11d1755b47b"
|
| 48 |
+
},
|
| 49 |
+
{
|
| 50 |
+
"path": "cards/README.ko.md",
|
| 51 |
+
"bytes": 11028,
|
| 52 |
+
"sha256": "f4f6b672714a2c9628c20e6b36bff8347809ef6d069239fcfea0d341a5c79d9d"
|
| 53 |
+
},
|
| 54 |
+
{
|
| 55 |
+
"path": "cards/README.pt-BR.md",
|
| 56 |
+
"bytes": 10462,
|
| 57 |
+
"sha256": "a182321cef7696ffb7f55f39e66377cedd6d182ed685765ade56b4ea00f55230"
|
| 58 |
+
},
|
| 59 |
+
{
|
| 60 |
+
"path": "cards/README.ru.md",
|
| 61 |
+
"bytes": 12482,
|
| 62 |
+
"sha256": "2c5cf4422c0e736d1686adf3770ac4cc0ddeb24e485120b55f10e40cd9376ca4"
|
| 63 |
+
},
|
| 64 |
+
{
|
| 65 |
+
"path": "cards/README.zh-CN.md",
|
| 66 |
+
"bytes": 10091,
|
| 67 |
+
"sha256": "c43cbf507a98514363a9dc0be60ab6d04f8dff1384386efa49092141595d4588"
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"path": "cards/README.zh-TW.md",
|
| 71 |
+
"bytes": 10088,
|
| 72 |
+
"sha256": "1c6a9833748af64e9a7e8c65d435abb84aa6a7b5fa69dbf212ae1e308db1f79e"
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"path": "documentation.json",
|
| 76 |
+
"bytes": 1315,
|
| 77 |
+
"sha256": "91542f5d2d39dddb9df041f51c6bea02fa4cde5166a1e37fb045064b3cb31c23"
|
| 78 |
+
}
|
| 79 |
+
]
|
| 80 |
+
}
|
README.md
CHANGED
|
@@ -25,71 +25,224 @@ task_categories:
|
|
| 25 |
- text-generation
|
| 26 |
---
|
| 27 |
|
| 28 |
-
|
|
|
|
|
|
|
|
|
|
| 29 |
|
| 30 |
## 한국어
|
| 31 |
|
| 32 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
|
| 34 |
-
원본
|
| 35 |
-
기존 정제본 2,148,459문서, 학습 2,125,938/검증 22,521문서입니다.
|
| 36 |
-
pathcosmos는 추가 정제·분리·패키징 제작자이며 원문 제작자는 원 출처를 따릅니다.
|
| 37 |
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
|
| 42 |
-
출처 복원 상태: `repo_path_row_preserved_git_commit_unknown`. 복원하지 않은 URL이나 원 Git commit을 확인된 것으로 표시하지 않습니다.
|
| 43 |
-
원본 라이선스의 범위와 미완료 고지는 [source.json](source.json), [LICENSE-SCOPE.md](LICENSE-SCOPE.md)를 확인하세요.
|
| 44 |
|
| 45 |
## English
|
| 46 |
|
| 47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
-
|
| 50 |
-
Original selection: 64 evenly spread shards out of 1126. Locally curated documents: 2,148,459.
|
| 51 |
-
pathcosmos performed additional curation and packaging; upstream authors retain attribution.
|
| 52 |
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
|
|
|
|
|
|
|
|
|
| 57 |
|
| 58 |
-
## Canonical split reader
|
| 59 |
|
| 60 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 61 |
|
| 62 |
```python
|
| 63 |
-
from huggingface_hub import
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
for document in download_and_iter(repo, revision=revision, split="train", max_documents=10):
|
| 68 |
-
print(document["doc_id"], document["text"][:120])
|
| 69 |
```
|
| 70 |
|
| 71 |
-
Use the pinned v1.0.0 tag or the commit in the main release index. `tokens/train/` and `tokens/validation/`
|
| 72 |
-
are little-endian uint16, EOS=1, using tokenizer SHA
|
| 73 |
-
`050f9b43cf964f0c3b29ae443a1c2eae46cb2ee0022cc010bcfea3d26acf1706`.
|
| 74 |
-
Different tokenizers require retokenizing text. Token offsets include the appended EOS.
|
| 75 |
|
| 76 |
-
|
| 77 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 78 |
|
| 79 |
-
|
| 80 |
-
## Repository links / 저장소 링크
|
| 81 |
|
| 82 |
-
[
|
|
|
|
| 83 |
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
- text-generation
|
| 26 |
---
|
| 27 |
|
| 28 |
+
|
| 29 |
+
# korean-pretrain-code
|
| 30 |
+
|
| 31 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 32 |
|
| 33 |
## 한국어
|
| 34 |
|
| 35 |
+
### 개요
|
| 36 |
+
|
| 37 |
+
이 저장소는 아래에 표시한 한 소스의 정제·분리 자료를 제공합니다. 동일 소스의 전체 파일은 통합 corpus의 해당 폴더에서도 받을 수 있습니다. 데이터의 고정 버전과 문서의 갱신 버전은 구분합니다.
|
| 38 |
+
|
| 39 |
+
문서 번역 언어는 학습 데이터에 그 언어가 포함된다는 뜻이 아닙니다. 주 자연언어는 한국어·영어이며 웹·수학·코드에 다른 언어가 섞일 수 있습니다.
|
| 40 |
+
|
| 41 |
+
### 구성과 규모
|
| 42 |
+
|
| 43 |
+
| | |
|
| 44 |
+
|---|---:|
|
| 45 |
+
| 소스 | github_code_v1 |
|
| 46 |
+
| 문서 수 | 2,148,459 |
|
| 47 |
+
| train 문서 | 2,125,938 |
|
| 48 |
+
| eval 문서 | 22,521 |
|
| 49 |
+
| 텍스트 bytes | 3,584,084,955 |
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
| | |
|
| 53 |
+
|---|---:|
|
| 54 |
+
| tokens | 4,006,737,196 |
|
| 55 |
+
| eval tokens | 42,420,481 |
|
| 56 |
+
| revision | v1.0.0 |
|
| 57 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 58 |
+
| upstream license | other |
|
| 59 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
### 처리와 분리
|
| 63 |
+
|
| 64 |
+
웹은 공백 정리, 길이·반복·언어 휴리스틱, 일부 개인정보 패턴 마스킹을 거칩니다. 초기 샤드 내부 hash와 최종 웹 NFC SHA256의 기준은 다릅니다. 코드 정제는 대소문자·들여쓰기·원문 Unicode를 보존하고 경로·라이선스 표기·생성 코드·일부 비밀정보 패턴을 검사합니다. 동일 본문·웹 NFC 동등 본문·코드 저장소의 연결 그룹을 train/eval로 분리합니다.
|
| 65 |
+
|
| 66 |
+
### 사용법
|
| 67 |
+
|
| 68 |
+
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.
|
| 69 |
+
|
| 70 |
+
```bash
|
| 71 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 72 |
+
```
|
| 73 |
+
|
| 74 |
+
```python
|
| 75 |
+
from huggingface_hub import snapshot_download
|
| 76 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 77 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 78 |
+
local_dir='korean-pretrain-code', token=False)
|
| 79 |
+
```
|
| 80 |
+
|
| 81 |
+
|
| 82 |
+
```python
|
| 83 |
+
from pathlib import Path
|
| 84 |
+
import subprocess, sys
|
| 85 |
+
from huggingface_hub import snapshot_download
|
| 86 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 87 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 88 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 89 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 90 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 91 |
+
```
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 95 |
+
|
| 96 |
+
### 모델과 학습
|
| 97 |
+
|
| 98 |
+
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.
|
| 99 |
+
|
| 100 |
+
### 최적화·벤치마크
|
| 101 |
+
|
| 102 |
+
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.
|
| 103 |
+
|
| 104 |
+
### 검증
|
| 105 |
+
|
| 106 |
+
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.
|
| 107 |
+
|
| 108 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 109 |
+
|
| 110 |
+
### 출처와 라이선스
|
| 111 |
|
| 112 |
+
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.
|
|
|
|
|
|
|
| 113 |
|
| 114 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 115 |
+
|
| 116 |
+
### 한계
|
| 117 |
+
|
| 118 |
+
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.
|
| 119 |
+
|
| 120 |
+
### 관련 저장소
|
| 121 |
+
|
| 122 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.ko.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 123 |
+
[GH / 사용법](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / 최적화·벤치마크](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / 검증](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 124 |
+
|
| 125 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.ko.md)
|
| 126 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.ko.md)
|
| 127 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.ko.md)
|
| 128 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.ko.md)
|
| 129 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.ko.md)
|
| 130 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.ko.md)
|
| 131 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.ko.md)
|
| 132 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md)
|
| 133 |
+
|
| 134 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 135 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 136 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 137 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 138 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 139 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 140 |
|
|
|
|
|
|
|
| 141 |
|
| 142 |
## English
|
| 143 |
|
| 144 |
+
### Overview
|
| 145 |
+
|
| 146 |
+
This repository provides the cleaned, split material for the single source identified below. The same source files are also available in its directory within the unified corpus. Pinned data versions and documentation revisions are distinct.
|
| 147 |
+
|
| 148 |
+
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.
|
| 149 |
|
| 150 |
+
### Contents and scale
|
|
|
|
|
|
|
| 151 |
|
| 152 |
+
| | |
|
| 153 |
+
|---|---:|
|
| 154 |
+
| Source | github_code_v1 |
|
| 155 |
+
| Documents | 2,148,459 |
|
| 156 |
+
| Train documents | 2,125,938 |
|
| 157 |
+
| Evaluation documents | 22,521 |
|
| 158 |
+
| Text bytes | 3,584,084,955 |
|
| 159 |
|
|
|
|
| 160 |
|
| 161 |
+
| | |
|
| 162 |
+
|---|---:|
|
| 163 |
+
| tokens | 4,006,737,196 |
|
| 164 |
+
| eval tokens | 42,420,481 |
|
| 165 |
+
| revision | v1.0.0 |
|
| 166 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 167 |
+
| upstream license | other |
|
| 168 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 169 |
+
|
| 170 |
+
|
| 171 |
+
### Processing and splits
|
| 172 |
+
|
| 173 |
+
Web processing applies whitespace cleanup, length/repetition/language heuristics and selected PII-pattern masking. Initial within-shard hashes differ from final web NFC SHA256. Code filtering preserves case, indentation and original Unicode while checking paths, license labels, generated code and selected secret patterns. Connected groups of identical bodies, NFC-equivalent web bodies and code repositories are assigned to train/eval.
|
| 174 |
+
|
| 175 |
+
### Usage
|
| 176 |
+
|
| 177 |
+
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.
|
| 178 |
+
|
| 179 |
+
```bash
|
| 180 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 181 |
+
```
|
| 182 |
|
| 183 |
```python
|
| 184 |
+
from huggingface_hub import snapshot_download
|
| 185 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 186 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 187 |
+
local_dir='korean-pretrain-code', token=False)
|
|
|
|
|
|
|
| 188 |
```
|
| 189 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 190 |
|
| 191 |
+
```python
|
| 192 |
+
from pathlib import Path
|
| 193 |
+
import subprocess, sys
|
| 194 |
+
from huggingface_hub import snapshot_download
|
| 195 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 196 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 197 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 198 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 199 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 200 |
+
```
|
| 201 |
+
|
| 202 |
+
|
| 203 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 204 |
+
|
| 205 |
+
### Model and training
|
| 206 |
+
|
| 207 |
+
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.
|
| 208 |
+
|
| 209 |
+
### Optimization and benchmarks
|
| 210 |
+
|
| 211 |
+
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.
|
| 212 |
+
|
| 213 |
+
### Validation
|
| 214 |
+
|
| 215 |
+
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.
|
| 216 |
+
|
| 217 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 218 |
+
|
| 219 |
+
### Sources and licensing
|
| 220 |
+
|
| 221 |
+
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.
|
| 222 |
+
|
| 223 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 224 |
+
|
| 225 |
+
### Limitations
|
| 226 |
+
|
| 227 |
+
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.
|
| 228 |
|
| 229 |
+
### Related repositories
|
|
|
|
| 230 |
|
| 231 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.en.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 232 |
+
[GH / Usage](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / Optimization and benchmarks](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / Validation](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 233 |
|
| 234 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.en.md)
|
| 235 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.en.md)
|
| 236 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.en.md)
|
| 237 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.en.md)
|
| 238 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.en.md)
|
| 239 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.en.md)
|
| 240 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.en.md)
|
| 241 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md)
|
| 242 |
|
| 243 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 244 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 245 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 246 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 247 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 248 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
cards/README.ar.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — العربية
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## نظرة عامة
|
| 6 |
+
|
| 7 |
+
يوفر هذا المستودع البيانات المنظفة والمقسمة للمصدر الواحد المحدد أدناه. توجد الملفات نفسها في مجلده داخل المجموعة الموحدة. إصدار البيانات المثبت ومراجعة التوثيق أمران منفصلان.
|
| 8 |
+
|
| 9 |
+
لغة الترجمة لا تعني وجود بيانات تدريب بتلك اللغة. اللغتان الأساسيتان الكورية والإنجليزية؛ قد تحتوي الويب والرياضيات والشيفرة لغات أخرى.
|
| 10 |
+
|
| 11 |
+
## المحتوى والحجم
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| المصدر | github_code_v1 |
|
| 16 |
+
| الوثائق | 2,148,459 |
|
| 17 |
+
| وثائق التدريب | 2,125,938 |
|
| 18 |
+
| وثائق التقييم | 22,521 |
|
| 19 |
+
| بايتات النص | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## المعالجة والتقسيم
|
| 33 |
+
|
| 34 |
+
تشمل معالجة الويب تنظيف الفراغات وقواعد الطول/التكرار/اللغة وإخفاء أنماط محددة من المعلومات الشخصية. تختلف بصمة الجزء الأولية عن NFC SHA256 النهائي للويب. تحافظ تصفية الشيفرة على حالة الأحرف والمسافات البادئة وUnicode الأصلي، وتفحص المسارات وتسميات الترخيص والشيفرة المولدة وبعض أنماط الأسرار. توزع المجموعات المتصلة للنصوص المتطابقة ونصوص الويب المكافئة في NFC ومستودعات الشيفرة على train/eval.
|
| 35 |
+
|
| 36 |
+
## الاستخدام
|
| 37 |
+
|
| 38 |
+
نفذ الأوامر من جذر المستودع. استخدم مجلدات إخراج وتنزيل جديدة وتحقق أولاً من النطاق والحجم وrevision المثبت. لا ينزل مثال CPU الاصطناعي بيانات خارجية. اختر الملفات اللازمة من الإصدارات المنفصلة وتحقق من الحجم/SHA256 باستخدام CHECKSUMS.json أو manifests المصادر.
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## النموذج والتدريب
|
| 67 |
+
|
| 68 |
+
النموذج مفكك ترميز كثيف بـPyTorch يستخدم RMSNorm وRoPE وGQA وSwiGLU. مسار3B الفعلي يستخدم FSDP2 و8GPU وeager:سياق4096،micro batch4،دفعة عالمية2,097,152tokens،ذروة LR3e-4 وهدف57,000step. مثال CPU الصغير مختلف عن التدريب الموزع؛ الاستئناف الدقيق يتطلب model وoptimizer وRNG وعقود المدخلات والإعدادات.
|
| 69 |
+
|
| 70 |
+
## التحسين والقياسات
|
| 71 |
+
|
| 72 |
+
تقارن التجارب worker وthread وbatch وذاكرة SQLite وRAM والتخزين المحلي/NFS وeager/compile والدقة وذاكرة GPU. الوقت الكلي والإنتاجية وذروة الذاكرة وتكافؤ المحتوى والأرقام تدعم أسباب القبول أو الرفض:47 حملة و99 ملف JSON. السرعة لا تبرر فشلاً عددياً، والعينة الصغيرة لا تثبت الحل الأمثل على كامل البيانات.
|
| 73 |
+
|
| 74 |
+
## التحقق
|
| 75 |
+
|
| 76 |
+
اجتاز إصدار GH المصدر284 اختبار CPU مع9 متجاوزة؛ واجتازت بيئة GitHub CPU الجديدة258 مع35 متجاوزة. تحتاج الاختبارات المتجاوزة لسجلات التشغيل الأصلية أو نظام ملفات ثانٍ. يستخدم المثال الاصطناعي120وثيقة وtrain100/eval20 وvocab272 ونموذج108,864معلمة ويتحقق من استعادة model/optimizer بدقة. فحوص بصمات البيانات وتقسيمها وreader ليست إعادة إنتاج3B مستقلة كاملة.
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## المصادر والتراخيص
|
| 81 |
+
|
| 82 |
+
الشيفرة الأصلية للمشروع MIT والشرح الأصلي CC BY4.0. لا تحل هذه المنح محل حقوق المحتوى الأصلي. راجع source.json وLICENSE-SCOPE.md ووثائق المصدر للشروط والنسب والإشعارات والمسائل المعلقة. لا ندعي حرية إعادة التوزيع التجاري غير المقيد لكل نص أصلي.
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## القيود
|
| 87 |
+
|
| 88 |
+
لا نضمن إزالة التشابه العالمي أو جميع المعلومات الشخصية/الأسرار أو خلو التقييم الرسمي من التلوث أو الجودة النهائية. كان التدريب الرئيسي جارياً عند اللقطة؛ لم تكتمل SFT أو تصدير النموذج أو المقاييس اللغوية الرسمية. لا توجد لقطة مستعادة لـ11مرجع شيفرة تاريخياً/9بصمات فريدة. تحتاج بعض أدوات التشغيل manifests الأصلية.
|
| 89 |
+
|
| 90 |
+
## المستودعات المرتبطة
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.ar.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / الاستخدام](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / التحسين والقياسات](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / التحقق](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.ar.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.ar.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.ar.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.ar.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.ar.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.ar.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.ar.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.de.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — Deutsch
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## Überblick
|
| 6 |
+
|
| 7 |
+
Dieses Repository enthält die bereinigten und aufgeteilten Daten der unten angegebenen einzelnen Quelle. Dieselben Dateien sind auch im zugehörigen Ordner des Gesamtkorpus verfügbar. Fixierte Datenversion und Dokumentationsrevision sind getrennt.
|
| 8 |
+
|
| 9 |
+
Übersetzungssprachen bedeuten keine entsprechenden Trainingsdaten. Hauptsprachen sind Koreanisch und Englisch; Web-, Mathematik- und Codeinhalte können weitere Sprachen enthalten.
|
| 10 |
+
|
| 11 |
+
## Inhalt und Umfang
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| Quelle | github_code_v1 |
|
| 16 |
+
| Dokumente | 2,148,459 |
|
| 17 |
+
| Trainingsdokumente | 2,125,938 |
|
| 18 |
+
| Evaluationsdokumente | 22,521 |
|
| 19 |
+
| Textbytes | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## Verarbeitung und Aufteilung
|
| 33 |
+
|
| 34 |
+
Webverarbeitung bereinigt Leerraum, nutzt Längen-/Wiederholungs-/Sprachheuristiken und maskiert ausgewählte personenbezogene Muster. Der anfängliche shardinterne Hash unterscheidet sich vom abschließenden NFC-SHA256. Code behält Großschreibung, Einrückung und ursprüngliches Unicode; Pfade, Lizenzangaben, generierter Code und ausgewählte Geheimnismuster werden geprüft. Verbundene Gruppen gleicher Texte, NFC-äquivalenter Webtexte und Code-Repositories werden train/eval zugeteilt.
|
| 35 |
+
|
| 36 |
+
## Verwendung
|
| 37 |
+
|
| 38 |
+
Die Befehle im Repository-Hauptverzeichnis ausführen. Neue Ausgabe-/Downloadordner verwenden und Umfang, Größe sowie fixierte Revision prüfen. Das synthetische CPU-Beispiel lädt keine externen Daten. Benötigte Dateien aus separaten Veröffentlichungen wählen und Größe/SHA256 anhand CHECKSUMS.json oder Quellmanifests prüfen.
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## Modell und Training
|
| 67 |
+
|
| 68 |
+
Der reine PyTorch-Dense-Decoder verwendet RMSNorm, RoPE, GQA und SwiGLU. Der reale3B-Pfad nutzt FSDP2,8GPUs und eager: Kontext4096, Microbatch4, global2,097,152Tokens, Spitzen-LR3e-4 und Ziel57,000Schritte. Das CPU-Microbeispiel unterscheidet sich vom verteilten Training. Exakte Wiederaufnahme verlangt Modell, Optimizer, RNG und Eingabe-/Konfigurationsverträge.
|
| 69 |
+
|
| 70 |
+
## Optimierung und Benchmarks
|
| 71 |
+
|
| 72 |
+
Verglichen werden Worker, Threads, Batches, SQLite-Cache, RAM, lokal/NFS, eager/compile, Präzision und GPU-Speicher. Gesamtzeit, Durchsatz, Speicherspitze und Inhalts-/numerische Gleichwertigkeit begründen Annahme oder Ablehnung:47 Kampagnen,99 Ergebnis-JSONs. Geschwindigkeit rechtfertigt keinen numerisch fehlerhaften Kandidaten; kleine Stichproben beweisen kein Gesamtoptimum.
|
| 73 |
+
|
| 74 |
+
## Validierung
|
| 75 |
+
|
| 76 |
+
Die GH-Quellversion bestand284 CPU-Tests bei9 Überspringungen; frische GitHub-CPU-CI bestand258 bei35 Überspringungen. Gründe sind fehlende Originalprotokolle oder ein zweites Dateisystem. Der synthetische Quickstart nutzt120Dokumente, train100/eval20, Vokabular272 und108,864Parameter und prüft exakte Modell-/Optimizer-Wiederaufnahme. Datenhashes, Splits und Readerprüfungen sind keine vollständige unabhängige3B-Reproduktion.
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## Quellen und Lizenzen
|
| 81 |
+
|
| 82 |
+
Originaler Projektcode steht unter MIT, eigener Erklärungstext unter CC BY4.0. Dies ersetzt nicht die Rechte an Quelldaten. source.json, LICENSE-SCOPE.md und Herkunftsdokumente nennen Bedingungen, Zuschreibungen, Hinweise und offene Punkte. Eine unbeschränkte kommerzielle Weitergabe aller Originaltexte wird nicht behauptet.
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## Grenzen
|
| 87 |
+
|
| 88 |
+
Globale Fast-Duplikatentfernung, vollständige PII-/Geheimnisentfernung, kontaminationsfreie Bewertung und finale Qualität werden nicht garantiert. Haupttraining lief beim Snapshot; SFT, Export und offizielle Sprachbenchmarks sind nicht abgeschlossen.11 historische Codeverweise/9einzigartige Hashes haben keinen wiedergefundenen Snapshot. Einige Betriebswerkzeuge benötigen Originalmanifests.
|
| 89 |
+
|
| 90 |
+
## Verwandte Repositories
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.de.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / Verwendung](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / Optimierung und Benchmarks](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / Validierung](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.de.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.de.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.de.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.de.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.de.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.de.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.de.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.en.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — English
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## Overview
|
| 6 |
+
|
| 7 |
+
This repository provides the cleaned, split material for the single source identified below. The same source files are also available in its directory within the unified corpus. Pinned data versions and documentation revisions are distinct.
|
| 8 |
+
|
| 9 |
+
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.
|
| 10 |
+
|
| 11 |
+
## Contents and scale
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| Source | github_code_v1 |
|
| 16 |
+
| Documents | 2,148,459 |
|
| 17 |
+
| Train documents | 2,125,938 |
|
| 18 |
+
| Evaluation documents | 22,521 |
|
| 19 |
+
| Text bytes | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## Processing and splits
|
| 33 |
+
|
| 34 |
+
Web processing applies whitespace cleanup, length/repetition/language heuristics and selected PII-pattern masking. Initial within-shard hashes differ from final web NFC SHA256. Code filtering preserves case, indentation and original Unicode while checking paths, license labels, generated code and selected secret patterns. Connected groups of identical bodies, NFC-equivalent web bodies and code repositories are assigned to train/eval.
|
| 35 |
+
|
| 36 |
+
## Usage
|
| 37 |
+
|
| 38 |
+
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.
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## Model and training
|
| 67 |
+
|
| 68 |
+
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.
|
| 69 |
+
|
| 70 |
+
## Optimization and benchmarks
|
| 71 |
+
|
| 72 |
+
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.
|
| 73 |
+
|
| 74 |
+
## Validation
|
| 75 |
+
|
| 76 |
+
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.
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## Sources and licensing
|
| 81 |
+
|
| 82 |
+
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.
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## Limitations
|
| 87 |
+
|
| 88 |
+
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.
|
| 89 |
+
|
| 90 |
+
## Related repositories
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.en.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / Usage](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / Optimization and benchmarks](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / Validation](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.en.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.en.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.en.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.en.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.en.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.en.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.en.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.es.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — Español
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## Descripción
|
| 6 |
+
|
| 7 |
+
Este repositorio ofrece los datos limpiados y divididos de la única fuente indicada abajo. Los mismos archivos están en su carpeta del corpus unificado. La versión fijada de datos y la revisión documental son distintas.
|
| 8 |
+
|
| 9 |
+
El idioma de traducción no implica datos de entrenamiento en ese idioma. Predominan coreano e inglés; web, matemáticas y código pueden contener otros.
|
| 10 |
+
|
| 11 |
+
## Contenido y escala
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| Fuente | github_code_v1 |
|
| 16 |
+
| Documentos | 2,148,459 |
|
| 17 |
+
| Documentos train | 2,125,938 |
|
| 18 |
+
| Documentos evaluación | 22,521 |
|
| 19 |
+
| Bytes texto | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## Procesamiento y división
|
| 33 |
+
|
| 34 |
+
La web pasa por limpieza de espacios, heurísticas de longitud/repetición/idioma y enmascaramiento de algunos patrones personales. El hash inicial por fragmento difiere del SHA256 NFC final. El código conserva mayúsculas, sangría y Unicode original; se filtran rutas, etiquetas de licencia, código generado y algunos secretos. Grupos conectados de texto idéntico, web equivalente en NFC y repositorios de código se asignan a train/eval.
|
| 35 |
+
|
| 36 |
+
## Uso
|
| 37 |
+
|
| 38 |
+
Ejecutar desde la raíz del repositorio. Usar directorios nuevos y comprobar alcance, tamaño y revision fijada. El ejemplo CPU sintético no descarga datos externos. Elegir los archivos necesarios de las publicaciones separadas y verificar tamaño/SHA256 con CHECKSUMS.json o manifests.
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## Modelo y entrenamiento
|
| 67 |
+
|
| 68 |
+
El decodificador denso PyTorch emplea RMSNorm, RoPE, GQA y SwiGLU. La ruta3B real usa FSDP2,8GPU y eager: contexto4096, micro batch4, batch global2,097,152tokens, LR máximo3e-4 y objetivo57,000pasos. El ejemplo CPU micro difiere del entrenamiento distribuido; la reanudación exacta exige modelo, optimizer, RNG y contratos de entrada/configuración.
|
| 69 |
+
|
| 70 |
+
## Optimización y mediciones
|
| 71 |
+
|
| 72 |
+
Se comparan workers, threads, batch, caché SQLite, RAM, local/NFS, eager/compile, precisión y memoria GPU. Tiempo total, rendimiento, pico de memoria y equivalencia de contenido/números sustentan cada adopción o rechazo:47 campañas y99 JSON. Ser más rápido no justifica fallos numéricos y una muestra pequeña no prueba el óptimo global.
|
| 73 |
+
|
| 74 |
+
## Validación
|
| 75 |
+
|
| 76 |
+
La versión fuente GH pasó284 pruebas CPU con9 omitidas; GitHub CPU nuevo pasó258 con35 omitidas. Se necesitan registros originales o un segundo sistema de archivos para las omitidas. El quickstart sintético usa120documentos, train100/eval20, vocab272 y108,864parámetros, verificando reanudación exacta de modelo/optimizer. Hash, división y reader de datos no equivalen a una reproducción3B independiente completa.
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## Fuentes y licencias
|
| 81 |
+
|
| 82 |
+
El código original es MIT y la explicación original CC BY4.0. No sustituyen los derechos del contenido fuente. Consultar source.json, LICENSE-SCOPE.md y procedencia para condiciones, atribución, avisos y asuntos pendientes. No se afirma redistribución comercial ilimitada de todos los textos.
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## Limitaciones
|
| 87 |
+
|
| 88 |
+
No se garantizan deduplicación aproximada global, eliminación total de PII/secretos, evaluación sin contaminación ni calidad final. El entrenamiento principal seguía activo en el snapshot; SFT, exportación y pruebas lingüísticas oficiales no están completados.11 referencias históricas/9hash únicos carecen de snapshot recuperado. Algunas herramientas requieren manifests originales.
|
| 89 |
+
|
| 90 |
+
## Repositorios relacionados
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.es.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / Uso](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / Optimización y mediciones](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / Validación](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.es.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.es.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.es.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.es.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.es.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.es.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.es.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.fr.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — Français
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## Présentation
|
| 6 |
+
|
| 7 |
+
Ce dépôt fournit les données nettoyées et séparées de l’unique source indiquée ci-dessous. Les mêmes fichiers figurent aussi dans son dossier du corpus unifié. La version figée des données et la révision documentaire sont distinctes.
|
| 8 |
+
|
| 9 |
+
La langue de traduction n’indique pas la présence de données d’entraînement dans cette langue. Les langues principales sont le coréen et l’anglais ; web, mathématiques et code peuvent en contenir d’autres.
|
| 10 |
+
|
| 11 |
+
## Contenu et volume
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| Source | github_code_v1 |
|
| 16 |
+
| Documents | 2,148,459 |
|
| 17 |
+
| Documents train | 2,125,938 |
|
| 18 |
+
| Documents évaluation | 22,521 |
|
| 19 |
+
| Octets texte | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## Traitement et séparation
|
| 33 |
+
|
| 34 |
+
Le web subit nettoyage des espaces, heuristiques de longueur/répétition/langue et masquage de certains motifs de données personnelles. Le hash initial interne aux fragments diffère du SHA256 NFC final. Le code conserve casse, indentation et Unicode original ; chemins, licences déclarées, code généré et certains secrets sont filtrés. Les groupes connectés de textes identiques, de textes web équivalents en NFC et de dépôts de code sont répartis en train/eval.
|
| 35 |
+
|
| 36 |
+
## Utilisation
|
| 37 |
+
|
| 38 |
+
Exécuter les commandes à la racine du dépôt. Utiliser de nouveaux dossiers et vérifier périmètre, taille et revision figée. L’exemple CPU synthétique ne télécharge pas de données externes. Choisir les fichiers nécessaires dans les distributions séparées et vérifier taille/SHA256 avec CHECKSUMS.json ou les manifests.
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## Modèle et entraînement
|
| 67 |
+
|
| 68 |
+
Le décodeur dense PyTorch utilise RMSNorm, RoPE, GQA et SwiGLU. Le parcours3B réel emploie FSDP2,8GPU et eager : contexte4096, micro batch4, batch global2,097,152tokens, pic LR3e-4, cible57,000steps. La démonstration CPU micro diffère du distribué. Une reprise exacte exige modèle, optimizer, RNG et contrats des entrées/configurations.
|
| 69 |
+
|
| 70 |
+
## Optimisation et mesures
|
| 71 |
+
|
| 72 |
+
Les essais comparent workers, threads, batches, cache SQLite, RAM, stockage local/NFS, eager/compile, précision et mémoire GPU. Temps total, débit, pic mémoire et équivalence de contenu/numérique accompagnent les décisions d’adoption ou de rejet :47 campagnes et99 JSON. Une accélération ne valide pas un candidat numériquement incorrect ; un petit échantillon ne prouve pas l’optimum global.
|
| 73 |
+
|
| 74 |
+
## Vérification
|
| 75 |
+
|
| 76 |
+
La version source GH compte284 tests CPU réussis/9 ignorés ; le nouvel environnement CPU GitHub en compte258/35. Les exclusions nécessitent des traces originales ou un second système de fichiers. Le quickstart synthétique utilise120documents, train100/eval20, vocab272 et108,864paramètres et vérifie la reprise exacte du modèle/optimizer. Hash, séparation et reader des données ne prouvent pas une reproduction3B indépendante complète.
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## Sources et licences
|
| 81 |
+
|
| 82 |
+
Le code original du projet est MIT, le texte explicatif original CC BY4.0. Ces droits ne remplacent pas ceux des contenus sources. Consulter source.json, LICENSE-SCOPE.md et les documents de provenance pour conditions, attributions, notices et points non résolus. Aucune redistribution commerciale illimitée de chaque texte original n’est affirmée.
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## Limites
|
| 87 |
+
|
| 88 |
+
La déduplication approximative globale, l’élimination totale des données personnelles/secrets, l’absence de contamination et la qualité finale ne sont pas garanties. L’entraînement principal était en cours au snapshot ; SFT, export et évaluations linguistiques officielles restent inachevés.11 références de code historiques/9hash uniques n’ont pas de snapshot retrouvé. Certains outils exigent les manifests originaux.
|
| 89 |
+
|
| 90 |
+
## Dépôts associés
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.fr.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / Utilisation](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / Optimisation et mesures](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / Vérification](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.fr.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.fr.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.fr.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.fr.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.fr.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.fr.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.fr.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.he.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — עברית
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## סקירה
|
| 6 |
+
|
| 7 |
+
מאגר זה מציע את הנתונים המנוקים והמחולקים של המקור היחיד המצוין להלן. אותם קבצים זמינים בתיקייה המתאימה בקורפוס המאוחד. גרסת הנתונים המקובעת ועדכון התיעוד נפרדים.
|
| 8 |
+
|
| 9 |
+
שפת התרגום אינה מעידה על נתוני אימון באותה שפה. השפות העיקריות הן קוריאנית ואנגלית;רשת,מתמטיקה וקוד עשויים להכיל שפות נוספות.
|
| 10 |
+
|
| 11 |
+
## תוכן והיקף
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| מקור | github_code_v1 |
|
| 16 |
+
| מסמכים | 2,148,459 |
|
| 17 |
+
| מסמכי אימון | 2,125,938 |
|
| 18 |
+
| מסמכי הערכה | 22,521 |
|
| 19 |
+
| בתי טקסט | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## עיבוד וחלוקה
|
| 33 |
+
|
| 34 |
+
עיבוד הרשת כולל ניקוי רווחים,כללי אורך/חזרה/שפה ומיסוך דפוסים מסוימים של מידע אישי. הגיבוב הראשוני בתוך מקטע שונה מNFC SHA256 הסופי. סינון הקוד שומר אותיות גדולות/קטנות,הזחה וUnicode מקורי,ובודק נתיבים,תוויות רישיון,קוד שנוצר ודפוסי סודות מסוימים. קבוצות מחוברות של טקסט זהה,טקסט רשת שקול בNFC ומאגרי קוד מחולקות לtrain/eval.
|
| 35 |
+
|
| 36 |
+
## שימוש
|
| 37 |
+
|
| 38 |
+
הריצו את הפקודות משורש המאגר. השתמשו בתיקיות פלט/הורדה חדשות ובדקו היקף,גודל וrevision מקובע. דוגמת CPU הסינתטית אינה מורידה נתונים חיצוניים. בחרו קבצים נחוצים מההפצות הנפרדות ואמתו גודל/SHA256 מול CHECKSUMS.json או manifests.
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## מודל ואימון
|
| 67 |
+
|
| 68 |
+
המודל הוא מפענח צפוף בPyTorch עם RMSNorm,RoPE,GQA וSwiGLU. מסלול3B האמיתי משתמש בFSDP2,8GPU וeager:הקשר4096,micro batch4,אצווה כוללת2,097,152tokens,שיא LR3e-4 ויעד57,000steps. דוגמת CPU micro שונה מאימון מבוזר. המשך מדויק דורש model,optimizer,RNG וחוזי קלט/תצורה.
|
| 69 |
+
|
| 70 |
+
## מיטוב ומדידות
|
| 71 |
+
|
| 72 |
+
הניסויים משווים worker,thread,batch,מטמון SQLite,RAM,אחסון מקומי/NFS,eager/compile,דיוק וזיכרון GPU. זמן כולל,תפוקה,שיא זיכרון ושקילות תוכן/מספרים מלווים סיבות לאימוץ או דחייה:47 מסעות ניסוי ו99 קובצי JSON. מהירות אינה מצדיקה כשל מספרי,ומדגם קטן אינו מוכיח מיטוב לכל הנתונים.
|
| 73 |
+
|
| 74 |
+
## אימות
|
| 75 |
+
|
| 76 |
+
גרסת המקור בGH עברה284 בדיקות CPU עם9 דילוגים;סביבת GitHub CPU חדשה עברה258 עם35 דילוגים. הדילוגים דורשים רשומות ריצה מקוריות או מערכת קבצים שנייה. הדוגמה הסינתטית משתמשת ב120מסמכים,train100/eval20,vocab272 ומודל108,864פרמטרים ומאמתת שחזור מדויק של model/optimizer. בדיקות גיבוב,חלוקה וreader של נתונים אינן שחזור3B עצמאי מלא.
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## מקורות ורישיונות
|
| 81 |
+
|
| 82 |
+
קוד הפרויקט המקורי ברישיון MIT וההסבר המקורי CC BY4.0. אלו אינם מחליפים זכויות בתוכן המקורי. תנאים לפי מקור,ייחוס,הודעות וסוגיות פתוחות מופיעים בsource.json,LICENSE-SCOPE.md ומסמכי המקור. אין טענה להיתר מסחרי בלתי מוגבל לכל טקסט.
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## מגבלות
|
| 87 |
+
|
| 88 |
+
אין הבטחה להסרת כפילויות מקורבות גלובלית,להסרה מלאה של מידע אישי/סודות,להערכה רשמית נקייה מזיהום או לאיכות סופית. האימון הראשי נמשך בעת תמונת המצב;SFT,ייצוא ומדדי שפה רשמיים לא הושלמו.11הפניות קוד היסטוריות/9גיבובים ייחודיים ללא תמונה מקורית ששוחזרה. חלק מכלי התפעול דורשים manifests מקוריים.
|
| 89 |
+
|
| 90 |
+
## מאגרים קשורים
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.he.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / שימוש](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / מיטוב ומדידות](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / אימות](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.he.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.he.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.he.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.he.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.he.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.he.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.he.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.hi.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — हिन्दी
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## परिचय
|
| 6 |
+
|
| 7 |
+
यह रिपॉज़िटरी नीचे बताए एक स्रोत की साफ़ और विभाजित सामग्री देती है। यही फ़ाइलें एकीकृत कॉर्पस के संबंधित फ़ोल्डर में भी हैं। निश्चित डेटा संस्करण और दस्तावेज़ संशोधन अलग हैं।
|
| 8 |
+
|
| 9 |
+
अनुवाद की भाषा उस भाषा के प्रशिक्षण डेटा की उपस्थिति नहीं दर्शाती। मुख्य प्राकृतिक भाषाएँ कोरियाई और अंग्रेज़ी हैं; वेब,गणित और कोड में दूसरी भाषाएँ हो सकती हैं।
|
| 10 |
+
|
| 11 |
+
## सामग्री और आकार
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| स्रोत | github_code_v1 |
|
| 16 |
+
| दस्तावेज़ | 2,148,459 |
|
| 17 |
+
| प्रशिक्षण दस्तावेज़ | 2,125,938 |
|
| 18 |
+
| मूल्यांकन दस्तावेज़ | 22,521 |
|
| 19 |
+
| पाठ बाइट्स | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## प्रसंस्करण और विभाजन
|
| 33 |
+
|
| 34 |
+
वेब प्रसंस्करण में खाली स्थान सुधार, लंबाई/दोहराव/भाषा नियम और कुछ व्यक्तिगत पैटर्न छिपाना शामिल है। शुरुआती shard के भीतर hash अंतिम वेब NFC SHA256 से अलग है। कोड में अक्षरों का case, indentation और मूल Unicode सुरक्षित हैं; पथ, लाइसेंस लेबल, generated code और कुछ secret पैटर्न जाँचे जाते हैं। समान पाठ, NFC-समतुल्य वेब पाठ और code repository के जुड़े समूह train/eval में बाँटे जाते हैं।
|
| 35 |
+
|
| 36 |
+
## उपयोग
|
| 37 |
+
|
| 38 |
+
नीचे के आदेश repository root से चलाएँ। नए output/download फ़ोल्डर लें और पहले दायरा, आकार तथा निश्चित revision जाँचें। कृत्रिम CPU उदाहरण बाहरी डेटा डाउनलोड नहीं करता। अलग डेटा प्रकाशन से आवश्यक फ़ाइलें चुनें और CHECKSUMS.json या source manifests से आकार/SHA256 सत्यापित करें।
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## मॉडल और प्रशिक्षण
|
| 67 |
+
|
| 68 |
+
शुद्ध PyTorch Dense decoder में RMSNorm, RoPE, GQA और SwiGLU हैं। वास्तविक3B पथ FSDP2,8GPU और eager उपयोग करता है:context4096,micro batch4,global batch2,097,152tokens,peak LR3e-4,लक्ष्य57,000steps। CPU micro उदाहरण वितरित प्रशिक्षण से अलग है। सही पुनःआरंभ में model,optimizer,RNG और input/configuration अनुबंध चाहिए।
|
| 69 |
+
|
| 70 |
+
## अनुकूलन और मापन
|
| 71 |
+
|
| 72 |
+
worker,thread,batch,SQLite cache,RAM,local/NFS,eager/compile,precision और GPU memory की तुलना की गई। कुल समय, throughput,peak memory और सामग्री/संख्यात्मक समानता के साथ अपनाने या अस्वीकार करने के कारण दिए हैं:47 अभियान और99 परिणाम JSON। तेज़ होना गलत संख्यात्मक परिणाम को सही नहीं बनाता; छोटा नमूना पूरे डेटा का सर्वोत्तम विकल्प नहीं सिद्ध करता।
|
| 73 |
+
|
| 74 |
+
## सत्यापन
|
| 75 |
+
|
| 76 |
+
GH स्रोत संस्करण में284 स्वतंत्र CPU परीक्षण सफल और9 छोड़े गए; नए GitHub CPU वातावरण में258 सफल और35 छोड़े गए। छोड़े गए परीक्षण मूल रन रिकॉर्ड या दूसरे filesystem पर निर्भर हैं। कृत्रिम quickstart में120दस्तावेज़,train100/eval20,vocab272 और108,864पैरामीटर हैं तथा model/optimizer का सटीक पुनःआरंभ जाँचा जाता है। डेटा hash,split और reader की जाँच पूर्ण स्वतंत्र3B पुनरुत्पादन नहीं है।
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## स्रोत और लाइसेंस
|
| 81 |
+
|
| 82 |
+
परियोजना का मूल कोड MIT और मूल व्याख्या CC BY4.0 है। ये स्रोत सामग्री के अधिकारों का स्थान नहीं लेते। स्रोत की शर्तें, श्रेय, सूचनाएँ और अनसुलझे मुद्दे source.json,LICENSE-SCOPE.md तथा provenance दस्तावेज़ों में देखें। प्रत्येक मूल पाठ के असीमित व्यावसायिक पुनर्वितरण का दावा नहीं है।
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## सीमाएँ
|
| 87 |
+
|
| 88 |
+
वैश्विक लगभग-समान डुप्लिकेट हटाने,संपूर्ण PII/secret हटाने,दूषणरहित आधिकारिक मूल्यांकन या अंतिम गुणवत्ता की गारंटी नहीं है। प्रकाशन snapshot पर मुख्य प्रशिक्षण चल रहा था; SFT,model export और आधिकारिक भाषा benchmarks पूरे नहीं हैं।11ऐतिहासिक code references/9अद्वितीय hash का मूल snapshot नहीं मिला। कुछ संचालन उपकरणों को मूल run manifests चाहिए।
|
| 89 |
+
|
| 90 |
+
## संबंधित रिपॉज़िटरी
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.hi.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / उपयोग](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / अनुकूलन और मापन](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / सत्यापन](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.hi.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.hi.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.hi.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.hi.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.hi.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.hi.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.hi.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.ja.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — 日本語
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## 概要
|
| 6 |
+
|
| 7 |
+
このリポジトリは下表の単一ソースについて、整形・分割した資料を提供します。同じファイルは統合コーパスの対応フォルダーでも取得できます。固定データ版と説明文の改訂版を区別します。
|
| 8 |
+
|
| 9 |
+
翻訳言語はその言語の学習データを含むことを意味しません。主自然言語は韓国語・英語で、ウェブ・数学・コードには他言語が混在し得ます。
|
| 10 |
+
|
| 11 |
+
## 構成と規模
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| ソース | github_code_v1 |
|
| 16 |
+
| 文書数 | 2,148,459 |
|
| 17 |
+
| 学習文書 | 2,125,938 |
|
| 18 |
+
| 評価文書 | 22,521 |
|
| 19 |
+
| テキストbytes | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## 処理と分割
|
| 33 |
+
|
| 34 |
+
ウェブは空白整理、長さ・反復・言語のヒューリスティック、一部PIIパターンのマスキングを適用します。初期シャード内ハッシュと最終ウェブNFC SHA256は基準が異なります。コードは大文字小文字・字下げ・元Unicodeを保持し、パス・ライセンス表示・生成コード・一部秘密情報パターンを検査します。同一本文、NFC等価ウェブ、コードリポジトリの連結グループをtrain/evalに分けます。
|
| 35 |
+
|
| 36 |
+
## 使い方
|
| 37 |
+
|
| 38 |
+
以下はリポジトリルートで実行します。新しい出力/取得先を使い、範囲・容量・固定revisionを確認してください。合成CPU例は外部データを取得しません。実データは別配布から必要なファイルを選び、CHECKSUMS.jsonまたはsource manifestのサイズ/SHA256を検証します。
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## モデルと学習
|
| 67 |
+
|
| 68 |
+
モデルはRMSNorm・RoPE・GQA・SwiGLUを用いる純PyTorch Dense decoderです。実3B経路はFSDP2・8GPU・eager、文脈4096、micro batch4、global batch2,097,152tokens、peak LR3e-4、目標57,000stepです。CPU micro例は分散学習と異なり、正確な再開にはモデル・optimizer・RNG・入力/設定契約が必要です。
|
| 69 |
+
|
| 70 |
+
## 最適化とベンチマーク
|
| 71 |
+
|
| 72 |
+
worker/thread/batch、SQLite cache、RAM、local/NFS、eager/compile、精度、GPUメモリを比較しました。総時間・処理量・peakメモリ・内容/数値等価性と採用/不採用理由を記録し、47キャンペーン・99結果JSONを提供します。高速でも数値条件に失敗した候補を推奨せず、小標本の最適値を全体最適としません。
|
| 73 |
+
|
| 74 |
+
## 検証
|
| 75 |
+
|
| 76 |
+
GHソース版の独立CPU検査は284成功/9skip、新GitHub CPU環境は258成功/35skipです。skipは元運用記録または別ファイルシステムが必要な検査です。合成例は120文書・train100/eval20・vocab272・108,864パラメータを使いモデル/optimizerの正確な再開を確認します。データhash・分割・reader検査と完全な独立3B再現は別です。
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## 出典とライセンス
|
| 81 |
+
|
| 82 |
+
プロジェクト独自コードはMIT、独自説明はCC BY4.0です。元データの権利を置き換えません。source.json、LICENSE-SCOPE.md、出典文書でソース別条件・帰属・告知・未解決事項を確認してください。全原文の無制限な商用再配布を主張しません。
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## 制限
|
| 87 |
+
|
| 88 |
+
全域近似重複除去、個人情報/secret完全除去、公式評価の無汚染、最終品質は保証しません。公開時点で本学習は進行中で、SFT・モデルexport・公式言語benchmarkは未完了です。過去コード11参照/9固有hashは元snapshot未回収です。一部運用ツールには元実行manifestが必要です。
|
| 89 |
+
|
| 90 |
+
## 関連リポジトリ
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.ja.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / 使い方](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / 最適化とベンチマーク](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / 検証](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.ja.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.ja.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.ja.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.ja.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.ja.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.ja.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.ja.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.ko.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — 한국어
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## 개요
|
| 6 |
+
|
| 7 |
+
이 저장소는 아래에 표시한 한 소스의 정제·분리 자료를 제공합니다. 동일 소스의 전체 파일은 통합 corpus의 해당 폴더에서도 받을 수 있습니다. 데이터의 고정 버전과 문서의 갱신 버전은 구분합니다.
|
| 8 |
+
|
| 9 |
+
문서 번역 언어는 학습 데이터에 그 언어가 포함된다는 뜻이 아닙니다. 주 자연언어는 한국어·영어이며 웹·수학·코드에 다른 언어가 섞일 수 있습니다.
|
| 10 |
+
|
| 11 |
+
## 구성과 규모
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| 소스 | github_code_v1 |
|
| 16 |
+
| 문서 수 | 2,148,459 |
|
| 17 |
+
| train 문서 | 2,125,938 |
|
| 18 |
+
| eval 문서 | 22,521 |
|
| 19 |
+
| 텍스트 bytes | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## 처리와 분리
|
| 33 |
+
|
| 34 |
+
웹은 공백 정리, 길이·반복·언어 휴리스틱, 일부 개인정보 패턴 마스킹을 거칩니다. 초기 샤드 내부 hash와 최종 웹 NFC SHA256의 기준은 다릅니다. 코드 정제는 대소문자·들여쓰기·원문 Unicode를 보존하고 경로·라이선스 표기·생성 코드·일부 비밀정보 패턴을 검사합니다. 동일 본문·웹 NFC 동등 본문·코드 저장소의 연결 그룹을 train/eval로 분리합니다.
|
| 35 |
+
|
| 36 |
+
## 사용법
|
| 37 |
+
|
| 38 |
+
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.
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## 모델과 학습
|
| 67 |
+
|
| 68 |
+
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·입���/설정 계약 검증이 필요합니다.
|
| 69 |
+
|
| 70 |
+
## 최적화·벤치마크
|
| 71 |
+
|
| 72 |
+
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.
|
| 73 |
+
|
| 74 |
+
## 검증
|
| 75 |
+
|
| 76 |
+
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.
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## 출처와 라이선스
|
| 81 |
+
|
| 82 |
+
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.
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## 한계
|
| 87 |
+
|
| 88 |
+
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.
|
| 89 |
+
|
| 90 |
+
## 관련 저장소
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.ko.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / 사용법](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / 최적화·벤치마크](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / 검증](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.ko.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.ko.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.ko.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.ko.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.ko.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.ko.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.ko.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.pt-BR.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — Português (Brasil)
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## Visão geral
|
| 6 |
+
|
| 7 |
+
Este repositório fornece dados limpos e divididos da única fonte indicada abaixo. Os mesmos arquivos estão na pasta correspondente do corpus unificado. A versão fixada dos dados e a revisão da documentação são distintas.
|
| 8 |
+
|
| 9 |
+
O idioma da tradução não implica dados de treino nesse idioma. As línguas principais são coreano e inglês; web, matemática e código podem conter outras.
|
| 10 |
+
|
| 11 |
+
## Conteúdo e escala
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| Fonte | github_code_v1 |
|
| 16 |
+
| Documentos | 2,148,459 |
|
| 17 |
+
| Documentos treino | 2,125,938 |
|
| 18 |
+
| Documentos avaliação | 22,521 |
|
| 19 |
+
| Bytes texto | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## Processamento e divisão
|
| 33 |
+
|
| 34 |
+
A web passa por limpeza de espaços, heurísticas de tamanho/repetição/idioma e mascaramento de alguns padrões pessoais. O hash inicial por shard difere do SHA256 NFC final. Código preserva maiúsculas, indentação e Unicode original; verificam-se caminhos, rótulos de licença, código gerado e alguns segredos. Grupos conectados de textos iguais, web equivalente em NFC e repositórios de código são divididos em train/eval.
|
| 35 |
+
|
| 36 |
+
## Uso
|
| 37 |
+
|
| 38 |
+
Execute na raiz do repositório. Use novas pastas e confira escopo, tamanho e revision fixada. O exemplo CPU sintético não baixa dados externos. Selecione arquivos necessários nas distribuições separadas e confira tamanho/SHA256 em CHECKSUMS.json ou manifests.
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## Modelo e treinamento
|
| 67 |
+
|
| 68 |
+
O decodificador denso PyTorch usa RMSNorm, RoPE, GQA e SwiGLU. O caminho3B real usa FSDP2,8GPU e eager: contexto4096, micro batch4, batch global2,097,152tokens, pico LR3e-4 e meta57,000passos. O exemplo CPU micro difere do distribuído. A retomada exata exige modelo, optimizer, RNG e contratos de entrada/configuração.
|
| 69 |
+
|
| 70 |
+
## Otimização e medições
|
| 71 |
+
|
| 72 |
+
Comparam-se workers, threads, batch, cache SQLite, RAM, local/NFS, eager/compile, precisão e memória GPU. Tempo total, vazão, pico de memória e equivalência de conteúdo/números sustentam adoção ou rejei��ão:47 campanhas e99 JSON. Velocidade não compensa falha numérica, e uma amostra pequena não prova ótimo global.
|
| 73 |
+
|
| 74 |
+
## Validação
|
| 75 |
+
|
| 76 |
+
A edição GH passou284 testes CPU com9 ignorados; o novo GitHub CPU passou258 com35 ignorados. Os ignorados exigem registros originais ou outro sistema de arquivos. O quickstart sintético usa120documentos, train100/eval20, vocab272 e108,864parâmetros, verificando retomada exata de modelo/optimizer. Hash, divisão e reader de dados não equivalem à reprodução3B independente completa.
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## Fontes e licenças
|
| 81 |
+
|
| 82 |
+
O código original é MIT e o texto original CC BY4.0. Essas permissões não substituem direitos do conteúdo fonte. Consulte source.json, LICENSE-SCOPE.md e procedência para condições, atribuição, avisos e pendências. Não se afirma redistribuição comercial irrestrita de todo texto original.
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## Limitações
|
| 87 |
+
|
| 88 |
+
Não se garantem deduplicação aproximada global, remoção total de PII/segredos, avaliação sem contaminação ou qualidade final. O treino principal seguia ativo no snapshot; SFT, exportação e benchmarks linguísticos oficiais não estão concluídos.11 referências históricas/9hash únicos não possuem snapshot recuperado. Algumas ferramentas precisam de manifests originais.
|
| 89 |
+
|
| 90 |
+
## Repositórios relacionados
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.pt-BR.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / Uso](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / Otimização e medições](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / Validação](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.pt-BR.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.pt-BR.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.pt-BR.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.pt-BR.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.pt-BR.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.pt-BR.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.pt-BR.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.ru.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — Русский
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## Обзор
|
| 6 |
+
|
| 7 |
+
Репозиторий предоставляет очищенные и разделённые данные одного источника, указанного ниже. Те же файлы есть в соответствующей папке общего корпуса. Фиксированная версия данных и редакция документации различаются.
|
| 8 |
+
|
| 9 |
+
Язык перевода не означает наличие обучающих данных на нём. Основные языки —корейский и английский; веб, математика и код могут содержать другие.
|
| 10 |
+
|
| 11 |
+
## Состав и объём
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| Источник | github_code_v1 |
|
| 16 |
+
| Документы | 2,148,459 |
|
| 17 |
+
| Документы train | 2,125,938 |
|
| 18 |
+
| Документы оценки | 22,521 |
|
| 19 |
+
| Байты текста | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## Обработка и разделение
|
| 33 |
+
|
| 34 |
+
Веб-обработка включает очистку пробелов, эвристики длины/повторов/языка и маскирование отдельных персональных шаблонов. Начальный хеш внутри шарда отличается от итогового NFC SHA256. Код сохраняет регистр, отступы и исходный Unicode; проверяются пути, лицензии, сгенерированный код и отдельные секреты. Связные группы одинаковых текстов, NFC-эквивалентных веб-текстов и кодовых репозиториев делятся на train/eval.
|
| 35 |
+
|
| 36 |
+
## Использование
|
| 37 |
+
|
| 38 |
+
Команды выполняются из корня репозитория. Используйте новые папки, проверьте объём и фиксированную revision. Синтетический CPU-пример не загружает внешние данные. Выберите нужные файлы в отдельных публикациях и сверьте размер/SHA256 с CHECKSUMS.json или manifest.
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## Модель и обучение
|
| 67 |
+
|
| 68 |
+
Плотный декодер PyTorch использует RMSNorm, RoPE, GQA и SwiGLU. Реальный3B-путь: FSDP2,8GPU,eager, контекст4096, micro batch4, глобально2,097,152токена, пик LR3e-4, цель57,000шагов. CPU micro пример отличается от распределённого обучения. Точное продолжение требует модель, optimizer, RNG и контракты входов/настроек.
|
| 69 |
+
|
| 70 |
+
## Оптимизация и измерения
|
| 71 |
+
|
| 72 |
+
Сравниваются worker/thread/batch, кэш SQLite, RAM, local/NFS, eager/compile, точность и память GPU. Полное время, пропускная способность, пик памяти и эквивалентность содержимого/чисел обосновывают принятие или отказ:47 кампаний и99 JSON. Скорость не оправдывает численные ошибки; малая выборка не доказывает глобальный оптимум.
|
| 73 |
+
|
| 74 |
+
## Проверка
|
| 75 |
+
|
| 76 |
+
GH-исходники прошли284 CPU-теста при9 пропусках; новый GitHub CPU —258 при35 пропусках. Пропуски требуют оригинальные записи или вторую файловую систему. Синтетический пример:120документов, train100/eval20, vocab272,108,864параметра; проверяется точное восстановление модели/optimizer. Хеши, разбиение и reader данных не означают полную независимую репродукцию3B.
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## Источники и лицензии
|
| 81 |
+
|
| 82 |
+
Исходный код проекта —MIT, авторские пояснения —CC BY4.0. Это не заменяет права на исходное содержание. См.source.json, LICENSE-SCOPE.md и происхождение: условия, атрибуция, уведомления и нерешённые вопросы. Не заявляется неограниченное коммерческое распространение каждого исходного текста.
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## Ограничения
|
| 87 |
+
|
| 88 |
+
Не гарантируются глобальная приблизительная дедупликация, полное удаление PII/секретов, отсутствие загрязнения оценок и итоговое качество. На момент снимка обучение продолжалось; SFT, экспорт и официальные языковые тесты не завершены.11 исторических ссылок на код/9уникальных hash не имеют восстановленного снимка. Некоторым инструментам нужны оригинальные manifest.
|
| 89 |
+
|
| 90 |
+
## Связанные репозитории
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.ru.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / Использование](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / Оптимизация и измерения](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / Проверка](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.ru.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.ru.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.ru.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.ru.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.ru.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.ru.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.ru.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.zh-CN.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — 简体中文
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## 概览
|
| 6 |
+
|
| 7 |
+
此仓库提供下表指定单一来源的清洗与划分数据。同一来源文件也可在统一语料库对应目录获取。固定数据版本与文档修订版本分别标明。
|
| 8 |
+
|
| 9 |
+
文档翻译语言不代表训练数据覆盖该语言。主要自然语言为韩语和英语，网页、数学及代码中可能混有其他语言。
|
| 10 |
+
|
| 11 |
+
## 组成与规模
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| 来源 | github_code_v1 |
|
| 16 |
+
| 文档数 | 2,148,459 |
|
| 17 |
+
| 训练文档 | 2,125,938 |
|
| 18 |
+
| 评估文档 | 22,521 |
|
| 19 |
+
| 文本字节 | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## 处理与划分
|
| 33 |
+
|
| 34 |
+
网页处理包括空白整理、长度/重复/语言启发式和部分个人信息模式掩码。最初分片内哈希与最终网页NFC SHA256标准不同。代码过滤保留大小写、缩进及原始Unicode，并检查路径、许可标签、生成代码与部分密钥模式。相同正文、NFC等价网页及代码仓库的连通组被分配到train/eval。
|
| 35 |
+
|
| 36 |
+
## 使用方法
|
| 37 |
+
|
| 38 |
+
以下命令在仓库根目录执行。使用新输出/下载目录，先核对范围、容量和固定revision。合成CPU示例不下载外部数据。实际数据请从独立发布中选择必要文件，并依据CHECKSUMS.json或来源manifest验证大小/SHA256。
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## 模型与训练
|
| 67 |
+
|
| 68 |
+
模型采用纯PyTorch稠密解码器，包含RMSNorm、RoPE、GQA和SwiGLU。实际3B路径使用FSDP2、8GPU和eager：上下文4096、micro batch4、global batch2,097,152tokens、峰值LR3e-4、目标57,000step。CPU微型示例与分布式训练不同。精确恢复需要模型、optimizer、RNG及输入/配置契约。
|
| 69 |
+
|
| 70 |
+
## 优化与基准
|
| 71 |
+
|
| 72 |
+
实验比较worker、thread、batch、SQLite缓存、RAM、本地/NFS、eager/compile、精度和GPU内存，同时记录总时间、吞吐、峰值内存、内容/数值等价性与采纳或拒绝原因。提供47个实验组和99个结果JSON。速度更快不代表数值失败的候选可用，小样本最优也不代表全量最优。
|
| 73 |
+
|
| 74 |
+
## 验证
|
| 75 |
+
|
| 76 |
+
GH源码版独��CPU测试284通过、9跳过；新GitHub CPU环境258通过、35跳过。跳过项需要原始运行记录或第二文件系统。合成示例使用120文档、train100/eval20、vocab272和108,864参数模型，并验证模型/optimizer精确恢复。数据内容哈希、划分和reader验证不等于完整独立3B复现。
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## 来源与许可
|
| 81 |
+
|
| 82 |
+
项目原创代码为MIT，原创说明为CC BY4.0，不替代原始内容权利。请查看source.json、LICENSE-SCOPE.md与来源文档中的特定条件、署名、通知及未决事项。不宣称每篇原文均可无限制商业再分发。
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## 局限
|
| 87 |
+
|
| 88 |
+
不保证全局近似去重、彻底清除个人信息/密钥、官方评估无污染或最终模型质量。发布快照时主训练仍在进行；SFT、模型导出及官方语言基准尚未完成。11个历史代码引用/9个唯一hash缺少原始快照。一些运维工具需要原运行manifest。
|
| 89 |
+
|
| 90 |
+
## 相关仓库
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.zh-CN.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / 使用方法](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / 优化与基准](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / 验证](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.zh-CN.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.zh-CN.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.zh-CN.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.zh-CN.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.zh-CN.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.zh-CN.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.zh-CN.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
cards/README.zh-TW.md
ADDED
|
@@ -0,0 +1,111 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# korean-pretrain-code — 繁體中文
|
| 2 |
+
|
| 3 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
| 4 |
+
|
| 5 |
+
## 概覽
|
| 6 |
+
|
| 7 |
+
此儲存庫提供下表指定單一來源的清理與切分資料。同一來源檔案也可於統整語料庫對應目錄取得。固定資料版本與文件修訂版本分別標示。
|
| 8 |
+
|
| 9 |
+
文件翻譯語言不代表訓練資料涵蓋該語言。主要自然語言是韓語與英語，網頁、數學及程式碼可能混有其他語言。
|
| 10 |
+
|
| 11 |
+
## 組成與規模
|
| 12 |
+
|
| 13 |
+
| | |
|
| 14 |
+
|---|---:|
|
| 15 |
+
| 來源 | github_code_v1 |
|
| 16 |
+
| 文件數 | 2,148,459 |
|
| 17 |
+
| 訓練文件 | 2,125,938 |
|
| 18 |
+
| 評估文件 | 22,521 |
|
| 19 |
+
| 文字位元組 | 3,584,084,955 |
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
| | |
|
| 23 |
+
|---|---:|
|
| 24 |
+
| tokens | 4,006,737,196 |
|
| 25 |
+
| eval tokens | 42,420,481 |
|
| 26 |
+
| revision | v1.0.0 |
|
| 27 |
+
| upstream | [codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) |
|
| 28 |
+
| upstream license | other |
|
| 29 |
+
| provenance_status | `repo_path_row_preserved_git_commit_unknown` |
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
## 處理與切分
|
| 33 |
+
|
| 34 |
+
網頁處理包含空白整理、長度/重複/語言啟發式與部分個資模式遮罩。最初分片內雜湊和最終網頁NFC SHA256標準不同。程式碼過濾保留大小寫、縮排及原始Unicode，並檢查路徑、授權標籤、產生的程式碼與部分金鑰模式。相同正文、NFC等價網頁及程式碼儲存庫的連通群組分配至train/eval。
|
| 35 |
+
|
| 36 |
+
## 使用方法
|
| 37 |
+
|
| 38 |
+
下列命令在儲存庫根目錄執行。使用新的輸出/下載目錄，先確認範圍、容量及固定revision。合成CPU範例不下載外部資料。實際資料請由獨立發布選取必要檔案，依CHECKSUMS.json或來源manifest驗證大小/SHA256。
|
| 39 |
+
|
| 40 |
+
```bash
|
| 41 |
+
python -m pip install "pyarrow>=19,<20" "huggingface_hub>=0.34,<1" "tokenizers>=0.21,<0.23" "transformers==4.55.2" "requests>=2,<3"
|
| 42 |
+
```
|
| 43 |
+
|
| 44 |
+
```python
|
| 45 |
+
from huggingface_hub import snapshot_download
|
| 46 |
+
root = snapshot_download('pathcosmos/korean-pretrain-code', repo_type="dataset",
|
| 47 |
+
revision='v1.0.0', allow_patterns=['text/*.parquet', 'source.json', 'manifest.json', 'release.json', 'LICENSE-SCOPE.md'],
|
| 48 |
+
local_dir='korean-pretrain-code', token=False)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
```python
|
| 53 |
+
from pathlib import Path
|
| 54 |
+
import subprocess, sys
|
| 55 |
+
from huggingface_hub import snapshot_download
|
| 56 |
+
reader = snapshot_download("pathcosmos/korean-pretrain-corpus", repo_type="dataset", revision="v1.1.0",
|
| 57 |
+
allow_patterns=["examples/*"], local_dir="corpus-reader", token=False)
|
| 58 |
+
subprocess.run([sys.executable, str(Path(reader) / "examples/stream_documents.py"),
|
| 59 |
+
"--repo", "pathcosmos/korean-pretrain-corpus", "--revision", "v1.1.0", "--source", "github_code_v1",
|
| 60 |
+
"--split", "train", "--limit", "10"], check=True)
|
| 61 |
+
```
|
| 62 |
+
|
| 63 |
+
|
| 64 |
+
[examples/stream_documents.py](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/examples/stream_documents.py) · [manifest.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/manifest.json)
|
| 65 |
+
|
| 66 |
+
## 模型與訓練
|
| 67 |
+
|
| 68 |
+
模型為純PyTorch稠密解碼器，包含RMSNorm、RoPE、GQA及SwiGLU。實際3B路徑採FSDP2、8GPU與eager：上下文4096、micro batch4、global batch2,097,152tokens、峰值LR3e-4、目標57,000step。CPU微型範例不同於分散式訓練；精確復原需模型、optimizer、RNG與輸入/設定契約。
|
| 69 |
+
|
| 70 |
+
## 最佳化與基準
|
| 71 |
+
|
| 72 |
+
實驗比較worker、thread、batch、SQLite快取、RAM、本機/NFS、eager/compile、精度與GPU記憶體，同時記錄總時間、吞吐、峰值記憶體、內容/數值等價性及採用或拒絕原因。提供47組實驗與99個結果JSON。較快不能彌補數值條件失敗，小樣本最佳也不代表全量最佳。
|
| 73 |
+
|
| 74 |
+
## 驗證
|
| 75 |
+
|
| 76 |
+
GH原始碼版獨��CPU測試284通過、9略過；新GitHub CPU環境258通過、35略過。略過項需原始執行紀錄或第二檔案系統。合成範例使用120文件、train100/eval20、vocab272及108,864參數模型，驗證模型/optimizer精確復原。資料內容雜湊、切分與reader驗證不等於完整獨立3B重現。
|
| 77 |
+
|
| 78 |
+
[release.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/release.json) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json)
|
| 79 |
+
|
| 80 |
+
## 來源與授權
|
| 81 |
+
|
| 82 |
+
專案原創程式碼為MIT，原創說明為CC BY4.0，不取代原始內容權利。請查source.json、LICENSE-SCOPE.md與來源文件的個別條件、姓名標示、告知及未決事項。不宣稱每篇原文都可無限制商業再發布。
|
| 83 |
+
|
| 84 |
+
[codeparrot/github-code](https://huggingface.co/datasets/codeparrot/github-code/blob/b5661e6b17396364b2bcf8e68977b0d28e1ebd19/README.md) · [source.json](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/source.json) · [LICENSE-SCOPE.md](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/LICENSE-SCOPE.md)
|
| 85 |
+
|
| 86 |
+
## 限制
|
| 87 |
+
|
| 88 |
+
不保證全域近似去重、完全移除個資/金鑰、官方評估無汙染或最終模型品質。發布快照時主訓練仍進行中；SFT、模型匯出及官方語言基準尚未完成。11個歷史程式碼引用/9個唯一hash缺少原始快照。部分維運工具需原執行manifest。
|
| 89 |
+
|
| 90 |
+
## 相關儲存庫
|
| 91 |
+
|
| 92 |
+
[GitHub](https://github.com/pathcosmos/korean-dense-llm/blob/main/cards/README.zh-TW.md) · [README](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code)
|
| 93 |
+
[GH / 使用方法](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_GUIDE.md) · [GH / 最佳化與基準](https://github.com/pathcosmos/korean-dense-llm/blob/main/experiments/README.md) · [GH / 驗證](https://github.com/pathcosmos/korean-dense-llm/blob/main/docs/PUBLIC_VALIDATION.md)
|
| 94 |
+
|
| 95 |
+
- [korean-pretrain-corpus](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/blob/main/cards/README.zh-TW.md)
|
| 96 |
+
- [korean-pretrain-training-data](https://huggingface.co/datasets/pathcosmos/korean-pretrain-training-data/blob/main/cards/README.zh-TW.md)
|
| 97 |
+
- [korean-pretrain-wikipedia](https://huggingface.co/datasets/pathcosmos/korean-pretrain-wikipedia/blob/main/cards/README.zh-TW.md)
|
| 98 |
+
- [korean-pretrain-fineweb2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb2/blob/main/cards/README.zh-TW.md)
|
| 99 |
+
- [korean-pretrain-hplt2](https://huggingface.co/datasets/pathcosmos/korean-pretrain-hplt2/blob/main/cards/README.zh-TW.md)
|
| 100 |
+
- [korean-pretrain-fineweb-edu](https://huggingface.co/datasets/pathcosmos/korean-pretrain-fineweb-edu/blob/main/cards/README.zh-TW.md)
|
| 101 |
+
- [korean-pretrain-math](https://huggingface.co/datasets/pathcosmos/korean-pretrain-math/blob/main/cards/README.zh-TW.md)
|
| 102 |
+
- [korean-pretrain-code](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md)
|
| 103 |
+
|
| 104 |
+
- [wiki_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/wiki_ko)
|
| 105 |
+
- [fineweb2_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb2_ko)
|
| 106 |
+
- [hplt_ko/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/hplt_ko)
|
| 107 |
+
- [fineweb_edu/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/fineweb_edu)
|
| 108 |
+
- [owm/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/owm)
|
| 109 |
+
- [github_code_v1/](https://huggingface.co/datasets/pathcosmos/korean-pretrain-corpus/tree/main/github_code_v1)
|
| 110 |
+
|
| 111 |
+
[한국어](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ko.md) · [English](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.en.md) · [简体中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-CN.md) · [繁體中文](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.zh-TW.md) · [日本語](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ja.md) · [Français](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.fr.md) · [Deutsch](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.de.md) · [Español](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.es.md) · [Português (Brasil)](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.pt-BR.md) · [हिन्दी](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.hi.md) · [العربية](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ar.md) · [עברית](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.he.md) · [Русский](https://huggingface.co/datasets/pathcosmos/korean-pretrain-code/blob/main/cards/README.ru.md)
|
documentation.json
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"languages": {
|
| 3 |
+
"ko": "한국어",
|
| 4 |
+
"en": "English",
|
| 5 |
+
"zh-CN": "简体中文",
|
| 6 |
+
"zh-TW": "繁體中文",
|
| 7 |
+
"ja": "日本語",
|
| 8 |
+
"fr": "Français",
|
| 9 |
+
"de": "Deutsch",
|
| 10 |
+
"es": "Español",
|
| 11 |
+
"pt-BR": "Português (Brasil)",
|
| 12 |
+
"hi": "हिन्दी",
|
| 13 |
+
"ar": "العربية",
|
| 14 |
+
"he": "עברית",
|
| 15 |
+
"ru": "Русский"
|
| 16 |
+
},
|
| 17 |
+
"github": "https://github.com/pathcosmos/korean-dense-llm",
|
| 18 |
+
"hf_repositories": [
|
| 19 |
+
"pathcosmos/korean-pretrain-corpus",
|
| 20 |
+
"pathcosmos/korean-pretrain-training-data",
|
| 21 |
+
"pathcosmos/korean-pretrain-wikipedia",
|
| 22 |
+
"pathcosmos/korean-pretrain-fineweb2",
|
| 23 |
+
"pathcosmos/korean-pretrain-hplt2",
|
| 24 |
+
"pathcosmos/korean-pretrain-fineweb-edu",
|
| 25 |
+
"pathcosmos/korean-pretrain-math",
|
| 26 |
+
"pathcosmos/korean-pretrain-code"
|
| 27 |
+
],
|
| 28 |
+
"data_versions": {
|
| 29 |
+
"corpus": "v1.1.0",
|
| 30 |
+
"training": "v1.0.1",
|
| 31 |
+
"standalone": "v1.0.0"
|
| 32 |
+
},
|
| 33 |
+
"documentation_version": "v0.1.1",
|
| 34 |
+
"same_language_links_final": true,
|
| 35 |
+
"weights": {
|
| 36 |
+
"wiki_ko": 0.005,
|
| 37 |
+
"fineweb2_ko": 0.29464577095297395,
|
| 38 |
+
"hplt_ko": 0.20035422904702602,
|
| 39 |
+
"fineweb_edu": 0.3,
|
| 40 |
+
"owm": 0.1,
|
| 41 |
+
"github_code_v1": 0.1
|
| 42 |
+
},
|
| 43 |
+
"scope": "Documentation languages do not change dataset language metadata",
|
| 44 |
+
"repository": "pathcosmos/korean-pretrain-code",
|
| 45 |
+
"documentation_tag": "v1.0.1"
|
| 46 |
+
}
|