LFM2.5-KO-CPT-Full-LFMStyle-Shards-20260627

Source-separated LFM-style CPT shards: Korean Wiki, finance, legal raw/tasks/RAG/bar answers, and terminal ToolBench.

This dataset is part of the LFM2.5-8B-A1B-KO-SFT / Agentic SFT workflow.

• Main SFT model: https://huggingface.co/LLM-OS-Models/LFM2.5-8B-A1B-KO-SFT
• CPT base model: https://huggingface.co/LLM-OS-Models/LFM2.5-8B-A1B-KO-CPT-FULL
• Agentic follow-up model: https://huggingface.co/LLM-OS-Models/LFM2.5-8B-A1B-KO-Agentic-SFT
• SFT GitHub: https://github.com/gyunggyung/LFM25-KO-SFT
• CPT GitHub: https://github.com/gyunggyung/LFM25-KO-CPT

Source Attribution

• Source-separated CPT shards for Korean Wiki, Korean finance, Korean legal raw/task/RAG/bar-answer data, and terminal/tool-use traces.
• Shard filenames are preserved under data/ so consumers can inspect or reweight domains independently.

Additional public references:

• Liquid LFM base model: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B
• Liquid chat template docs: https://docs.liquid.ai/lfm/key-concepts/chat-template
• Liquid tool-use docs: https://docs.liquid.ai/lfm/key-concepts/tool-use
• Legalize-KR organization: https://github.com/legalize-kr
• KoTSQA v2.0: https://huggingface.co/datasets/etri-lirs/KoTSQA-v.2.0
• Korean dataset index reviewed for candidates: https://github.com/gyunggyung/LLM-Ko-Datasets

Notes

• Use this repo when you need source-level filtering, deduplication, or domain-ratio reconstruction.
• This is the best public artifact for auditing which CPT source files entered the KO-CPT run.

Summary

field	value
kind	raw_lfm_chat_jsonl
sample count	n/a
token count	n/a
max sequence / sample length	n/a
uploaded size bytes	26204095030

Format

• raw_lfm_chat_jsonl: JSONL rows with a text field containing LFM ChatML-like conversation text.
• prepared_tokenized: NumPy response-only SFT arrays built with the LFM tokenizer:
  • tokens.npy
  • epoch_0/inst_start.npy
  • epoch_0/inst_len.npy
  • epoch_0/resp_start.npy
  • epoch_0/resp_len.npy
  • tokenizer.json

Local Source Path

/home/work/.data/lfm2_ko_cpt/datasets/shards_full_lfmstyle_20260627

License And Usage Notes

This release republishes preprocessing artifacts used for the LFM2.5 Korean CPT/SFT workflow. Source components come from multiple public or locally prepared datasets, so downstream users should verify each upstream source license before redistribution or commercial use. Legal and finance examples are for model training/evaluation only and are not legal, financial, or investment advice.

Stats

{
  "path": "/home/work/.data/lfm2_ko_cpt/datasets/shards_full_lfmstyle_20260627",
  "size_bytes": 26204095030,
  "jsonl_files": [
    {
      "name": "000_kowiki_raw_full_20260524.jsonl",
      "size_bytes": 2055390454
    },
    {
      "name": "001_bcai_finance_kor_hrm_20260524.jsonl",
      "size_bytes": 5677401098
    },
    {
      "name": "002_BCAI-Finance-Kor-1862K.jsonl",
      "size_bytes": 5664370381
    },
    {
      "name": "003_korean_legal_raw_full_20260523.jsonl",
      "size_bytes": 1801038575
    },
    {
      "name": "004_korean_legal_tasks_full_20260524.jsonl",
      "size_bytes": 4130351212
    },
    {
      "name": "005_korean_admrule_precedent_raw_full_20260524.jsonl",
      "size_bytes": 1682358397
    },
    {
      "name": "006_ko_legal_source_agent_sft_20260621.jsonl",
      "size_bytes": 123458032
    },
    {
      "name": "007_ko_legal_rag_agent_sft_round15_v2.jsonl",
      "size_bytes": 14971282
    },
    {
      "name": "008_current_law_bar_json_answer_sft_20260621.jsonl",
      "size_bytes": 8900012
    },
    {
      "name": "009_lfm25_terminal_toolbench_hrm_turns_v1.jsonl",
      "size_bytes": 5045852445
    }
  ],
  "file_count": 20
}