Add files using upload-large-folder tool

README.md

@@ -0,0 +1,72 @@
+---
+license:
+- odc-by
+- cc-by-4.0
+language:
+- ko
+- en
+task_categories:
+- text-generation
+size_categories:
+- 10M<n<100M
+pretty_name: Perdix Pretrain Data
+tags:
+- pretraining
+- web
+- math
+- korean
+configs:
+- config_name: korean
+  data_files: fineweb2-kor/data/kor_Hang/train/*.parquet
+- config_name: web_en
+  data_files: dclm-baseline/**/*.parquet
+- config_name: math
+  data_files: finemath-4plus/finemath-4plus/*.parquet
+---
+
+# Perdix-Pretrain-Data
+
+[Perdix-1.1B-Base](https://huggingface.co/prismdata/Perdix-1.1B-Base) 사전학습에 쓴 원본 데이터를 그대로 모아 둔 것입니다. 새로 만든 데이터는 없고, 아래 세 공개 데이터셋에서 실제로 쓴 parquet 파일만 골라 복사했습니다. 어떤 파일을 어떤 비율로 썼는지 남겨 두려는 목적입니다.
+
+The exact subset of three public datasets (DCLM-baseline, FineWeb2 kor_Hang, FineMath 4+) used to pretrain Perdix-1.1B-Base. Files are unmodified copies; this repository only records which shards were used. 335 parquet files, about 89.7M documents, 178 GB.
+
+## 구성
+
+| 폴더 | 원 데이터셋 | 파일 | 문서 수 | 크기 | 토큰 수* | 라이선스 |
+|---|---|---|---|---|---|---|
+| `fineweb2-kor/` | [HuggingFaceFW/fineweb-2](https://huggingface.co/datasets/HuggingFaceFW/fineweb-2) `kor_Hang/train` 전부 | 25 | 60.9M | 105.8 GB | 45.6B | ODC-By |
+| `dclm-baseline/` | [mlfoundations/dclm-baseline-1.0-parquet](https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0-parquet) 일부 | 246 | 22.1M | 53.9 GB | 29.7B | CC-BY-4.0 |
+| `finemath-4plus/` | [HuggingFaceTB/finemath](https://huggingface.co/datasets/HuggingFaceTB/finemath) `finemath-4plus` 전부 | 64 | 6.7M | 18.4 GB | 10.8B | ODC-By |
+
+\* Perdix 토크나이저(BPE 49,152)로 센 값. 문서 사이 `<|endoftext|>` 포함.
+
+DCLM-baseline은 전체가 수 TB라 일부만 썼습니다. global-shard 01~10 각각의 `local-shard_0_of_10`에서 앞쪽 파일만 골랐고(01~06은 shard 0~24, 07~10은 shard 0~23), 고르는 규칙은 `download_raw_data.sh`에 그대로 있습니다. FineWeb2 한국어와 FineMath 4+는 train 전부입니다.
+
+각 parquet의 컬럼은 원 데이터셋 그대로이며, 학습에는 `text` 컬럼만 썼습니다.
+
+## 학습에 어떻게 썼나
+
+- 세 소스를 합쳐 86B 토큰 중 30B 토큰을 학습했습니다. 각 소스에서 2,048토큰 창을 무작위로 뽑는 방식이라 같은 구간이 우연히 다시 뽑힐 수는 있지만, 전체를 한 바퀴 돌지는 않았습니다.
+- 믹싱 비율은 영어 65% / 한국어 30% / 수학 5%에서 시작해 25% / 50% / 25%로 서서히 바꿨습니다.
+- 토큰화·패킹 코드는 `pretrain_data.py`, 학습 코드는 `pretrain.py`로 [github.com/theprismdata/Perdix](https://github.com/theprismdata/Perdix)에 있습니다.
+
+## 받기
+
+```bash
+hf download prismdata/Perdix-Pretrain-Data --repo-type dataset --local-dir raw-data
+```
+
+원 저장소에서 같은 파일을 직접 받으려면 `download_raw_data.sh`를 쓰면 됩니다. 결과 폴더 구조가 같습니다.
+
+```bash
+DEST=raw-data ./download_raw_data.sh
+```
+
+## 라이선스
+
+이 저장소는 원 데이터셋의 라이선스를 그대로 따릅니다.
+
+- `fineweb2-kor/`, `finemath-4plus/`: [ODC-By 1.0](https://opendatacommons.org/licenses/by/1-0/). 원 데이터셋 카드에 따라 [Common Crawl 이용약관](https://commoncrawl.org/terms-of-use)도 함께 적용됩니다.
+- `dclm-baseline/`: [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/).
+
+출처를 밝힐 때는 이 저장소가 아니라 원 데이터셋(FineWeb2, DCLM-baseline, FineMath)을 인용해 주세요. 원 데이터셋에서 특정 문서의 삭제 요청이 처리되면 이 저장소에도 반영하겠습니다.

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000000_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56003b694ee713c1be4068c501dc4990acfbb088ccf0b91ad891047e809b5dc1
+size 144473349

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000001_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:111ce6fe80ffc13692b2d10753530b2cc79ae4ebb958747c9284ee14d3ac5b96
+size 223934950

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000002_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c4cd8bca0f6b9d606784e844e087bf93f8dcae476b92e30173fb1174e114aa4
+size 241088104

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000003_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70b278c25081e09f63ff8b6c0a61ffa65d4af039e00022735a436081022599be
+size 243928978

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000004_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63a155e5efd8ac2e57542090d8e8ecd9ab5c80c8074d6483d64e22ffc2e26e91
+size 218407296

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000005_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc0a2dfb1388da84de5dc5c128e928ab4edbb3ebdbc8bed3d4c8b4843cf7d8f3
+size 199821264

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000006_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14c14db7f330c0d8e4834e8dc9916f78333c50b44b62276c6be66756390fc1a0
+size 222621751

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000007_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42e2d72476af521834defe66879579df6c19835113a0e90cc91cde855785ef5a
+size 204563840

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000008_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7729e42be9fb16e4f1ddeb10b905a3fb8814cc3744afd2a8a4fbb7d0df5a66e
+size 262945697

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000009_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7cae2b93f39187c2787e1bc5b7617f6feb372e50107ebaba70862a37f10bf819
+size 191858630

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000010_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bafa7fc450854e7cb5a4f57520cb1f17d920b460d933b12f6a73910b4a270a85
+size 204726229

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000011_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4dba36fdc253b3269c70012f5d6a7f782f0c70cf540508a1998100944a1411cf
+size 262690675

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000012_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d718ce07b0ce4f31f7ce2f005950083d6fe62f48912e7f560444a139a0639eff
+size 220287875

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000013_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea608f02cedacc4c34fa1c6201f4a9f80c77b1fed008a2cfb883e5f27224430b
+size 217584560

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000014_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a8dc80d0b75f8425a6d09ae869f8d22c01a6ed9a6548ca925a502d38552b40c
+size 228797127

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000015_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:974ed2b0e0c99738dc8cec552a9dad56b19a91de248c17ccf1142e25c9e30020
+size 215018611

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000016_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59be41261d07495fcd5198d732dbd2b3c68733528f27ba9c8dc20f272fea7c98
+size 132068954

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000017_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95a916422ed5aa522316a229f6401b556e5c4adb85b8e1086316445ee50a5cf5
+size 210343647

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000018_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0061bbe71230fd483e6f5e423f42ee64acc169d37d7f2fd5e725d3e09d05a10
+size 222433373

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000019_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ba76da4d75129b12fb96fb8f1ddb18a8c6452aecde79949e012ff7a53ee89a7
+size 205615427

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000020_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac393895c36cb559b58a310aa8bdf24f402b98577afa0f43a87a66b3885cf9e3
+size 293195333

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000021_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:409dadc9258a6b523870d4e2efa401139ef6753e52f24988d75ddde302e6a0d1
+size 227672186

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000022_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:311192948f8f805d71fc75c51a6b9b1d5b65b1c0b1a83491f1a207d5fb29fdd1
+size 208116259

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000023_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d022ee3ae64af8078f42becad9436d89cafd228e6ebfddd78bc9532bb01d0ec
+size 247066775

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_01_of_10/local-shard_0_of_10/shard_00000024_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a30a8cefb38c9430e3475877a64c688b70d9e7d671d92ac09688a3c1de9cd3a3
+size 287610023

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000000_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40deb9e9b2d95360328957e25b8892fe0e3049ab3c4542c947920616235764e2
+size 136445768

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000001_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:831cd6e547ca29c47e64542423f4d7b8dca5fc9019c05d34a2cd1a6f1d8ec349
+size 153054327

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000002_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56a378ca94b329d9129d8b3c8dbe2c9a7f2a01f3ab546c0ff59b5636fc018234
+size 181756399

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000003_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af916ce09a2bbef045e8ed2603f356cd79ec311f0ebc1a976e3e70dd353807c1
+size 194442902

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000004_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b62393eadc65a1d5a5acaf8823adee795c5299020041e23403250fcb92e4a798
+size 201451854

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000005_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd3a462892a0fa661451ce943680f4f7da43671aa0b3ccc1f960e22a39b7b9af
+size 196024218

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000006_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d58bc2d125e8ad375bc6d8094bb60fafaab5ce5092bcf5a51270d5ea6007e421
+size 228503751

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000007_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77fc2657d0b551631357e0b0f56c84133b55fda1314e01414a4b8993a51d5b6f
+size 240760141

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000008_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:261fbee87017cb929a0af7ec8e65382c254a6e4185daccdaa7d386c46e63217e
+size 226933438

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000009_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fc6de994241215eb848e954ce46bf54f509653556d492a93a79dd5a9444560b
+size 219392896

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000010_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17985008e8f1c219e1166bc7f5761e7878e33bad7d37134fb0c6fbdef6e9f114
+size 214927837

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000011_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59b5b6a63a5af457dabc8775376d7b49a73378432a9746812ae94c98989d33fa
+size 222760987

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000012_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ec0a7e921570e984aae5a6c5ba7bd5965de92c590619c792a1f592fe4be1c91
+size 221205159

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000013_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d4dfa06e13cf93750ac812d3902f792d17f3ddd3c9985fb7edb100c19673416e
+size 209277827

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000014_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0b18f1211c7564ca62a8cc543de66134776ada90c727893057acd11d23655a2b
+size 233492661

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000015_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76acec6d030cb47b480338d1de06e039a7a31417f3df9134fe80fa5168d94ba2
+size 207760305

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000016_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80acb1be38e9ed1096e705f4f3cb43f4959523c8c821b313aeea0f046d241862
+size 251739918

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000017_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fe3f60f80e6b8f68dc5198c69a5aeb507d029703f7c1b210466cbe94c9c2df9
+size 222119611

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000018_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3a7c8b129626454316be16490b9a75835046a165779028162025374ad42a24a
+size 221487805

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000019_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2cc515f78ddad1a566d532c03006a2cd33ce752ff9efd77c0caa3adc3db62647
+size 227572639

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000020_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31bca5dbae1eacd14bbb23f97c7fd9f1f5dca0ec27894ad7c8b3d17d1c527dcd
+size 205613159

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000021_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6b5dc65ea066b8d5de97cc0613126e1f66c35c45cb22e3ac40e06f2cf4dd9687
+size 213414214

dclm-baseline/filtered/OH_eli5_vs_rw_v2_bigram_200k_train/fasttext_openhermes_reddit_eli5_vs_rw_v2_bigram_200k_train/processed_data/global-shard_02_of_10/local-shard_0_of_10/shard_00000022_processed.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fab0587d8a758bc1024ea235f9aa072a6c17100ba61d34b27b45317f251bc66
+size 246379190

download_raw_data.sh

@@ -0,0 +1,62 @@
+#!/bin/bash
+# Perdix 사전학습에 쓴 원본 데이터(parquet)를 허깅 페이스 원 저장소에서 받는다.
+# (prismdata/Perdix-Pretrain-Data 에 같은 파일이 그대로 올라가 있으므로
+#  그쪽을 통째로 받아도 된다: hf download prismdata/Perdix-Pretrain-Data --repo-type dataset --local-dir raw-data)
+#
+#   수학    FineMath 4+        64개 파일   18.4GB  -> raw-data/finemath-4plus
+#   영어 웹  DCLM-baseline     246개 파일   53.9GB  -> raw-data/dclm-baseline
+#   한국어 웹 FineWeb2 kor_Hang  25개 파일  105.8GB  -> raw-data/fineweb2-kor
+#
+# 실행:
+#   ./download_raw_data.sh              # 전부 받기 (끊기면 다시 실행하면 이어서 받음)
+#   ./download_raw_data.sh --dry-run    # 받지 않고 받을 파일만 확인
+#   DEST=/다른/경로 ./download_raw_data.sh
+set -euo pipefail
+
+DEST="${DEST:-$(pwd)/raw-data}"
+EXTRA=("$@")
+
+# 외장하드 등에 받을 때 마운트가 빠져 있으면 엉뚱한 곳에 받으므로 상위 폴더부터 확인한다.
+if [ ! -d "$(dirname "$DEST")" ]; then
+    echo "폴더가 없습니다: $(dirname "$DEST") (외장하드 연결 확인)" >&2
+    exit 1
+fi
+if ! command -v hf >/dev/null; then
+    echo "hf 명령이 없습니다: pip install -U huggingface_hub" >&2
+    exit 1
+fi
+mkdir -p "$DEST"
+
+download() {
+    local name="$1" repo="$2"
+    shift 2
+    echo "=== $name ($repo) ==="
+    hf download "$repo" --repo-type dataset --local-dir "$DEST/$name" "$@" ${EXTRA[@]+"${EXTRA[@]}"}
+}
+
+# 수학: finemath-4plus 폴더 전부
+download finemath-4plus HuggingFaceTB/finemath \
+    --include "finemath-4plus/*"
+
+# 영어 웹: 전체가 수 TB라 일부만 쓴다. global-shard 01~10 각각의 local-shard_0_of_10에서
+# 앞쪽 파일만(01~06은 shard 0~24, 07~10은 shard 0~23). packed/web_en의 246개와 같다.
+DCLM="local-shard_0_of_10/shard"
+download dclm-baseline mlfoundations/dclm-baseline-1.0-parquet \
+    --include "*global-shard_0[1-6]_of_10/${DCLM}_000000[01]?_processed.parquet" \
+    --include "*global-shard_0[1-6]_of_10/${DCLM}_0000002[0-4]_processed.parquet" \
+    --include "*global-shard_0[7-9]_of_10/${DCLM}_000000[01]?_processed.parquet" \
+    --include "*global-shard_0[7-9]_of_10/${DCLM}_0000002[0-3]_processed.parquet" \
+    --include "*global-shard_10_of_10/${DCLM}_000000[01]?_processed.parquet" \
+    --include "*global-shard_10_of_10/${DCLM}_0000002[0-3]_processed.parquet"
+
+# 한국어 웹: kor_Hang의 train 전부
+download fineweb2-kor HuggingFaceFW/fineweb-2 \
+    --include "data/kor_Hang/train/*"
+
+# 받은 파일 수 확인 (기대값: 64 / 246 / 25)
+echo "=== 결과 ==="
+for name in finemath-4plus dclm-baseline fineweb2-kor; do
+    n=$(find "$DEST/$name" -name "*.parquet" 2>/dev/null | wc -l | tr -d ' ')
+    size=$(du -sh "$DEST/$name" 2>/dev/null | cut -f1)
+    echo "$name: parquet ${n}개, ${size:-0}"
+done