mkd-minju/keural-v2-fluency

<div align="center">

Keural-v2 Fluency: A Curated Korean–English Conversational Corpus for LLM Fine-Tuning

</div>

<p align="center"> <a href="#1-introduction">Introduction</a> | <a href="#2-dataset-composition">Dataset Composition</a> (<a href="#23-source-revisions">Revisions</a>) | <a href="#3-data-collection--processing-methodology">Methodology</a> | <a href="#4-license-and-attribution">License</a> | <a href="#5-known-limitations">Limitations</a> </p>

Status: Private staging — not yet cleared for public release (pending §4 evaluation and second-party license audit).

1. Introduction

Keural-v2 Fluency is the "Area 1" component of the Keural-v2 Korean SFT training corpus, purpose-built for DeepSeek-V4-Flash-0731 fine-tuning. While the broader Keural-v2 program targets specialized capabilities such as tool-calling and chain-of-thought reasoning, this dataset addresses a distinct and easily-overlooked objective: preserving and improving the model's raw conversational fluency — the ability to produce natural, idiomatic Korean, without the "translationese" artifacts that plague machine-translated instruction data — while simultaneously guarding against degradation of the model's existing English conversational ability during Korean-focused fine-tuning.

To meet this goal without introducing the licensing and quality risks documented throughout this project's data-collection process (see §4 below), every candidate source was screened not only at the level of its top-level license tag, but down to its actual composition: several well-known, seemingly clean Korean instruction datasets were rejected after tracing their construction tables back to non-commercial, unlicensed, or machine-translated upstream components. The dataset below reflects only sources that survived this audit intact.

The corpus totals 300,000 examples, split 80/20 between Korean (240,000) and English (60,000), matching the project's overall fluency-objective ratio.

2. Dataset Composition

2.1 Korean Slice (240,000 examples, 80% of Area 1)

<div align="center">

</div>

jojo0217/korean_rlhf_dataset was evaluated as a candidate primary source but excluded after its construction table revealed that 91% of its rows trace to non-commercially-licensed (Stanford Alpaca, CC BY-NC 4.0), unlicensed (KoAlpaca v1.1), or machine-translated (Dolly-ko, OpenAssistant-ko) upstream content — none of which is curable by downstream rewriting.

2.2 English Balance Slice (60,000 examples, 20% of Area 1)

<div align="center">

</div>

This slice was deliberately restricted to human-authored data. Given how much of the rest of the Keural-v2 corpus is necessarily LLM-synthesized, the English balance slice is held to a stricter standard specifically to avoid compounding synthetic-data risk in the one category whose only job is to preserve an already-solid baseline capability.

2.3 Source Revisions

Exact commit/revision pinned at download time (2026-08-06), for reproducibility and future re-verification:

<div align="center">

</div>

3. Data Collection & Processing Methodology

1. 1.Each source was downloaded from its canonical location and exact-match deduplicated internally on its primary text field.
2. 2.Large, synthesis-heavy sources (keural-conversation-ko, IkJun1/korean-qa-dataset) were capped prior to pooling to prevent any single source from dominating the mix, preserving topical and stylistic diversity.
3. 3.All Korean-slice sources were pooled and cross-source deduplicated, then randomly sampled (seed=42) down to the 240,000-example target.
4. 4.OpenAssistant/oasst2 conversations were reconstructed by walking each valid English, non-deleted assistant reply back through its parent_id ancestry (rather than treating messages as independent rows), preserving genuine multi-turn structure; threads were capped at 8 turns.
5. 5.The English-slice supplementary sources (Taskmaster/MultiWOZ) were pooled with a per-source cap of 4,000, cross-deduplicated, and randomly sampled to close the exact gap remaining after Dolly and oasst2.
6. 6.Every record retains a _source and _license field; share-alike sources additionally carry _share_alike: true for downstream compliance tracking.

Final-format encoding (e.g., DeepSeek-V4-Flash-0731's native <｜User｜>/<｜Assistant｜>/<think> scheme) has intentionally not been applied at this stage — all records remain in structured role/content-style fields so that the correct target-model encoding can be applied uniformly across all five Keural-v2 data areas once collection is complete.

4. License and Attribution

This is a mixed-license dataset. Each record carries its own _source and _license field; verify the applicable license before any commercial use of a given slice.

<div align="center">

</div>

Share-alike notice. Rows derived from beomi/KoAlpaca-RealQA and databricks/databricks-dolly-15k carry _share_alike: true. Any redistribution or relicensing of this slice must satisfy CC BY-SA's attribution and same-license-propagation terms.

Gemma "Model Derivative" notice. Rows derived from keural-conversation-ko — the largest single component of the Korean slice — were generated by Google's Gemma (gemma-4-26B-A4B-it). Under Section 3.1 of the Gemma Terms of Use, a model fine-tuned on Gemma-generated synthetic outputs is itself considered a Gemma "Model Derivative." Anyone distributing a model fine-tuned on this dataset must: (1) include a notice stating "Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms," and (2) incorporate the Gemma Prohibited Use Policy into that model's own terms of use. This obligation is purely a downstream distribution/notice requirement — Google claims no ownership rights over Gemma-generated Outputs, so the dataset itself carries no copyright encumbrance.

This dataset inherits the licenses of its 13 upstream sources as listed above (Apache-2.0 / MIT / CC BY 4.0 / CC BY-SA 3.0 / CC BY-SA 4.0), mixed and processed in accordance with each source's attribution and share-alike terms. "Keural-v2 Fluency" is a dataset made by MKD.

5. Known Limitations

• —The generation models behind IkJun1/korean-qa-dataset and coastral/korean-writing-style-instruct are undisclosed by their original authors. This does not constitute a license violation (per the same disclosure standard already applied to glaiveai/glaive-function-calling-v2 in Area 2 of this project), but independent native-speaker quality review is recommended before final release.
• —The Taskmaster/MultiWOZ portion of the English slice consists of task-oriented (booking/reservation/inquiry) dialogue, a narrower register than Dolly/oasst2's open-domain conversation.
• —This dataset represents 300,000 of the project's overall 500,000-example target across five capability areas.
• —Target-model-specific chat-template encoding has not yet been applied (see §3); this happens in a unified pass across all five areas once collection is complete.

6. MKD Ownership Notice / 주식회사 MKD 소유권 고지

"Keural-v2 Fluency"는 주식회사 MKD(MKD Inc.)가 제작한 데이터셋입니다. 위 §3에서 설명한 수집 및 처리 작업(출처 심사, 중복 제거, 소스별 상한 설정, 교차 소스 샘플링, OpenAssistant 멀티턴 재구성)은 주식회사 MKD가 수행했으며, 원본 데이터는 위 §4에 명시된 각 원 저작자의 라이선스를 그대로 따릅니다.

"Keural-v2 Fluency" is a dataset produced by MKD Inc. (주식회사 MKD). The collection and processing work described in §3 above (source screening, deduplication, per-source capping, cross-source sampling, and OpenAssistant multi-turn reconstruction) was performed by MKD Inc.; underlying source data remains governed by the licenses of its original publishers as listed in §4.