KG-ZEROIN/zeroin-v8-r4-dataset

Zeroin Methodology v8-r4 QA

Training corpus used to fine-tune KG-ZEROIN/gpt-oss-20b-zeroin-v8-r4.

The corpus captures question/answer pairs derived from the Zeroin fund-evaluation methodology Korean domain document. It is organized as a Harmony-ready chat-messages dataset for supervised full fine-tuning of openai/gpt-oss-20b.

Released under CC BY-NC 4.0 — free for non-commercial research, evaluation, and educational use. See LICENSE and NOTICE.

Contents

File	Description
seed_v8_round4_section_all_full_ft.jsonl	2,638 Korean QA examples covering the entire document (round4, section_all)
v9p6_source.csv	Source CSV (236 Q/A groups, 15,240 expanded question variants) used to generate the JSONL
README.md	This dataset card
LICENSE	CC BY-NC 4.0 license text
NOTICE	Source attribution and derived-model references

Dataset structure

Each line of the JSONL is a single conversation record:

{
  "id": "zeroin.seed_v8_round4_section_all_full_ft_0001",
  "messages": [
    {"role": "system",    "content": "..."},
    {"role": "user",      "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "meta": {
    "dataset_version": "seed_v8_round4_section_all_full_ft",
    "round": "round4",
    "section": "section_all",
    "chapter": "1",
    "group_id": "group1",
    "question_variant_index": 1,
    "source_row_num": "1",
    "training_mode": "full_ft",
    "source_strategy": "csv_multiline_question_expansion"
  }
}

• The system role contains a fixed Korean-language domain instruction roughly equivalent to: "You are a domain assistant that answers based on the Zeroin fund-evaluation methodology."
• Each source Q/A group (group_id) is expanded into multiple semantically equivalent user-question paraphrases (question_variant_index) while keeping the assistant answer stable. This gives the model robustness to wording variation while staying anchored to the document-grounded answer.

Splits

Single train split with 2,638 records (~7.5 MB).

Validation and evaluation of the fine-tuned model were done via held-out prompt review reports, not via a separate split in this file.

Source

• Primary source document: a Korean Zeroin fund-evaluation methodology corpus.
• The upstream CSV (v8-r4 section-all.csv) was converted into messages-style JSONL through a multiline-question expansion strategy; see the originating project (scripts/ and llm_datasets/seed_v8/ in the source worktree) for the exact generation logic.

Intended use

• Reproducing or re-running the fine-tune of gpt-oss-20b that produced KG-ZEROIN/gpt-oss-20b-zeroin-v8-r4.
• Evaluating drift or regression of that fine-tune over time.
• Non-commercial research on Korean domain-specific full fine-tuning.

Out-of-scope use

• Commercial use of any kind (prohibited by CC BY-NC 4.0). Training, evaluating, or distributing commercial AI products on the basis of this dataset is not permitted under this license.

Licensing

Released under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International). Attribution to KG-ZEROIN/zeroin-v8-r4-dataset is required; commercial use is prohibited. Full text in LICENSE.

Citation

Zeroin Methodology v8-r4 QA dataset,
KG-ZEROIN/zeroin-v8-r4-dataset, 2026. (CC BY-NC 4.0)

사사 (Acknowledgment)

사사번호 (Project ID): RQT-25-090024

이 성과는 정부(과학기술정보통신부)의 재원으로 AI컴퓨팅자원활용기반강화(GPU 임차 지원) 사용자 지원사업의 지원을 받아 수행된 연구임.

This research was supported by the AI Computing Infrastructure Enhancement (GPU Rental Support) User Support Program funded by the Ministry of Science and ICT (MSIT), Republic of Korea.