Datasets:
Search is not available for this dataset
input_ids list | length uint16 |
|---|---|
[
124894,
124899,
5922,
207,
2857,
12807,
27637,
9246,
207,
110746,
3091,
27936,
4692,
9884,
26433,
43265,
5467,
89905,
20143,
9265,
69822,
17026,
22,
49890,
5467,
4889,
8206,
12624,
27936,
4692,
20,
13235,
14108,
12077,
4889,
20,
45539,
69892,... | 8,192 |
[
17723,
247,
3774,
60937,
107189,
20,
70393,
15727,
12023,
33307,
16079,
6501,
8008,
3756,
23968,
11222,
3091,
40717,
71022,
5467,
69769,
7620,
19143,
1957,
70402,
107791,
27829,
318,
8097,
12023,
3091,
13235,
6103,
16926,
12077,
107647,
21181,
... | 8,192 |
[
16,
40026,
2862,
4001,
70037,
32197,
11115,
20,
110204,
2235,
69246,
16380,
107381,
20,
108005,
109409,
32197,
6103,
20,
62244,
13846,
52835,
16,
25270,
44130,
16986,
17,
107542,
11222,
46278,
39837,
3756,
70165,
33392,
17462,
22,
31511,
1202... | 8,192 |
[
544,
132,
110,
110467,
38841,
70499,
5467,
32378,
58522,
3277,
7088,
1149,
243,
1578,
22,
8892,
7121,
20,
544,
132,
110,
110467,
38841,
70499,
1588,
70176,
1957,
70494,
4254,
26584,
31875,
8206,
2125,
55953,
5669,
49125,
7011,
46869,
2086,
... | 8,192 |
[1727,13657,28339,20,58171,15129,13521,13829,72793,5669,24198,24195,107816,70584,1588,9726,4692,1073(...TRUNCATED) | 8,192 |
[69901,72052,69813,72087,22,56460,56983,108548,3259,10174,13622,52770,3756,8625,44107,71790,2235,110(...TRUNCATED) | 8,192 |
[34365,36984,42141,6080,46143,318,1588,21185,229,1580,24,9237,6103,21649,28341,18122,3859,2857,3538,(...TRUNCATED) | 8,192 |
[108935,48447,4254,14989,11019,8008,8190,20,8794,26,16413,24807,32866,69811,42604,13249,36197,10733,(...TRUNCATED) | 8,192 |
[318,654,13736,229,916,6867,640,13362,33200,17457,6867,58019,2235,6784,8097,29845,121,3774,318,654,1(...TRUNCATED) | 8,192 |
[229,1163,6867,47412,28097,34,229,30,13736,229,1163,6867,207,18140,23334,108355,207,29,13736,229,249(...TRUNCATED) | 8,192 |
End of preview. Expand in Data Studio
LFM2.5-8B-A1B Korean CPT Data
Prepared Korean continued-pretraining data for LLM-OS-Models/LFM2.5-8B-A1B-KO-CPT-FULL.
Files
data/ko_cpt_mix_full_lfmstyle_20260627.jsonl: prepared full CPT corpus with one JSON object per line and atextfieldmetadata/ko_cpt_mix_full_lfmstyle_20260627.stats.json: corpus statisticsmetadata/ko_cpt_mix_full_lfmstyle_20260627.stats.json.full_report.json: per-source preprocessing reportmetadata/ko_cpt_sources_full_20260627.json: source configurationtokenized/tokenized_lfm25_8k_20260628/: packed 8192-token Parquet blocks will be uploaded after generation
Corpus Statistics
- Rows after global deduplication: 4,622,971
- Characters: 11,581,567,658
- Estimated tokens: 6,492,697,020
- Target tokenizer/model:
LiquidAI/LFM2.5-8B-A1B - Target sequence length: 8192
- Estimated packed training steps at effective batch 64: 12,384
- Tokenized export format: Parquet shards with
input_idsandlength
Format
Raw Korean wiki, legal, and finance documents are kept as plain completion text. Instruction, legal RAG, and terminal/tool-use examples are converted to LFM chat-style text using:
<|startoftext|><|im_start|><|im_end|>- roles:
system,user,assistant,tool
Each row:
{"text": "...", "source": "...", "category": "..."}
Source Mix
| Source | Rows |
|---|---|
kowiki_raw_full_20260524 |
611,403 |
bcai_finance_kor_hrm_20260524 |
1,861,531 |
korean_legal_raw_full_20260523 |
227,687 |
korean_legal_tasks_full_20260524 |
1,383,340 |
korean_admrule_precedent_raw_full_20260524 |
203,477 |
ko_legal_source_agent_sft_20260621 |
5,999 |
ko_legal_rag_agent_sft_round15_v2 |
749 |
current_law_bar_json_answer_sft_20260621 |
2,000 |
lfm25_terminal_toolbench_hrm_turns_v1 |
326,785 |
Legal Data Attribution
Legal-domain data is attributed to Legalize-KR and related Korean legal corpora.
Legalize-KR links:
- Organization: https://github.com/legalize-kr
- Korean statutes: https://github.com/legalize-kr/legalize-kr
- Korean court precedents: https://github.com/legalize-kr/precedent-kr
- Korean administrative rules: https://github.com/legalize-kr/admrule-kr
- Korean local ordinances: https://github.com/legalize-kr/ordinance-kr
- Collection/conversion pipeline: https://github.com/legalize-kr/legalize-pipeline
- Website: https://legalize.kr
- Public source: https://www.law.go.kr
Recommended citation:
- Statutes: include repository, Markdown path,
법령ID,법령MST, promulgation date, effective date, andlaw.go.krsource URL. - Precedents: include repository, Markdown path,
판례일련번호, court name, decision date, and case number. - Administrative rules: include repository, Markdown path, rule serial number or issuing number when available.
- Local ordinances: include repository, Markdown path,
자치법규ID,자치법규일련번호, promulgation date, promulgation number, and source URL. - Do not use only Git commit hashes as long-term identifiers because Legalize-KR notes that history may be reconstructed after parser or normalization improvements.
Korean Notes
이 데이터셋은 LiquidAI/LFM2.5-8B-A1B의 한국어 CPT를 위해 만든 공개용 준비 corpus입니다. 법률 데이터 출처는 Legalize-KR 및 관련 로컬 한국 법률 corpus를 명시합니다. Legalize-KR README 기준 법령/판례/행정규칙/자치법규 원문은 국가법령정보센터 OpenAPI에서 가져오며, 원문은 대한민국 정부 공공저작물로 안내되어 있습니다. 재현성을 위해서는 commit hash만이 아니라 법령ID, 판례일련번호, 자치법규ID, 날짜, law.go.kr URL을 함께 보존하는 것이 좋습니다.
Intended Use
This dataset is intended for continued pretraining and controlled evaluation of Korean language model behavior. It should not be used as legal advice, financial advice, or a substitute for authoritative legal texts.
- Downloads last month
- -