Hikaf143/keural-v2-fluency-v2

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

Introduction | Dataset Composition | Methodology | License | Limitations

Status: Private staging — full §3 processing pipeline complete (dedup → PII removal → Korean-purity filter → length check → ratio measurement → train/val/test split → schema normalization → target-model encoding). Pending §4 quantitative/qualitative evaluation and second-party license audit before any public release.

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. It preserves and improves the model's raw conversational fluency — natural, idiomatic Korean without "translationese" artifacts — while guarding against degradation of existing English conversational ability during Korean-focused fine-tuning.

This v2 release supersedes the earlier v1 staging snapshot: a full re-audit of the processing pipeline (2026-08-10) found and corrected several defects described in §3 and §5 below. The corpus now totals 290,636 examples (235,213 Korean / 55,423 English, ≈80.9%/19.1%, close to the ~80/20 target), each pre-encoded into DeepSeek-V4-Flash-0731's native training text format and split into train/val/test.

2. Dataset Composition

Sources are unchanged from v1 (see the original collection-time README for full per-source license/provenance/revision tables: mkd-chanwoo/keural-conversation-ko, IkJun1/korean-qa-dataset, coastral/korean-writing-style-instruct, CarrotAI/ko-instruction-dataset, beomi/KoAlpaca-RealQA for Korean; databricks/databricks-dolly-15k, OpenAssistant/oasst2, Taskmaster-1/2/3, MultiWOZ 2.2 for English). What changed in v2 is the processing, not the source list.

3. Data Collection & Processing Methodology

Full §3 pipeline applied and re-verified end-to-end (2026-08-10):

1. Deduplication: exact + near-duplicate (MinHash/LSH) removal, both intra-source and cross-source. 300,000 → 295,121 rows.
2. PII removal: regex-based masking (email/phone/card/RRN/IP) — but a full-instance audit found the naive first pass had a high false-positive rate specific to this category: IP-pattern matches were almost entirely coding/networking Q&A content (ping examples, router default-IP guides, Flask/Docker configs — none real personal IPs), and several CARD matches were plain numeric content (binary/hex conversions) with no "card" context. These patterns were disabled/gated (IP disabled entirely for this category; CARD applied only when "card"/"credit" appears in the same turn; RRN disabled for Korean since all matches were corporate registration numbers, not personal IDs — a different legal category entirely). Genuine email/phone matches were retained.
3. Korean purity filter: lang:"ko" rows checked for unexplained Hanja contamination — 4,470 rows removed.
4. Length/tokenizer check: real DeepSeek-V4-Flash-0731 tokenizer used. 5 Korean rows were found to be a degenerate generation-repetition bug from IkJun1/korean-qa-dataset (the same sentence repeated verbatim until ~74,500 characters) and removed — not truncated, since truncation would leave the repetition pattern intact.
5. Category/language ratio: measured directly (not assumed) — 80.9% ko / 19.1% en, accepted as within tolerance of the ~80/20 target.
6. Train/val/test split: performed with cross-split deduplication (0 leakage confirmed).
7. Schema normalization: all heterogeneous collection-time formats (flat instruction/response fields, ShareGPT-style, OpenAI-style) unified into schema/korean_sft_schema.md's common Turn structure.
8. Target-model encoding: DeepSeek's official encode_messages() implementation is not available in the public deepseek-ai/DeepSeek-V4-Flash-0731 HF repo (no chat_template, no custom code). Encoding was reimplemented from the tokenizer's own added_tokens (confirmed real, registered tokens: <｜User｜>, <｜Assistant｜>, <|EOT|>; fluency has no reasoning/tool-call content so <think> is unused here). During this step, two further defects were caught and fixed: (a) coastral's Vicuna-style role: "gpt" turns were not being mapped to assistant (encoder was raising on unrecognized roles rather than silently mis-encoding); (b) 1,912 Taskmaster1/2/3 dialogues ended on a trailing user utterance with no assistant reply to predict — these were trimmed to end on the last assistant turn, and 10 dialogues that had no assistant-preceding user turn at all (a loader artifact) were dropped entirely.

Post-encoding, the full token-length check was re-run across all 6 splits: 0 rows over the 32,768-token threshold.

4. License and Attribution

Unchanged from v1 — see the original README for the full per-source license table and the Gemma "Model Derivative" notice applicable to keural-conversation-ko-derived rows.

"Keural-v2 Fluency" is a dataset produced by 주식회사 MKD (MKD Inc.). The processing pipeline (deduplication, PII removal, filtering, splitting, and target-model encoding) described in §3 was performed by MKD Inc.; underlying source data retains the licenses of its original upstream publishers as listed above.

5. Known Limitations

• The DeepSeek-V4-Flash-0731 encoding used here is a reimplementation, not the vendor's official encode_messages() (which is not published). It was built from tokenizer-confirmed special tokens, but has not been validated against the vendor's actual reference implementation.
• Not yet evaluated per §4 (row/token distributions, diversity, contamination check against eval benchmarks) — pending before any public release.
• Represents 290,636 of the project's overall ~500,000-example target across six capability areas.