ENERZAiKR/huibeom_agentic-think-v2-react-hybrid-ko

agentic-think-v2-react-hybrid-ko

Korean localization of HBKenerzai/agentic-think-v2-react-hybrid: agentic tool-calling conversations where every natural-language surface — user messages, assistant reasoning, assistant replies, and user-derived argument strings — is rendered in Korean, while the API layer (function catalogs, parameter names, enum values, identifiers, tool observations) stays in English, matching real deployments.

181,894 rows after a Korean-purity pass (2026-10-04) that removed rows containing Chinese characters, translator/instruction leakage, single-turn register mixing, and emoji inside tool-call arguments. The assistant text keeps its original Markdown and emoji (an earlier spoken/TTS-flattened variant has been reverted at the maintainer's request). Same schema as the English corpus; the distractor-augmented tool catalogs (9–16 entries per episode) are inherited unchanged.

What is / isn't translated

Korean	kept as-is (English)
user messages	tool catalogs (tools)
assistant reasoning_content	parameter names, enum values, ids, codes
assistant replies	tool observations (role: tool)
user-composed argument strings (message bodies, notes)	schema-constrained argument values

Argument language follows how a Korean speaker actually talks: titles, brands and technical terms stay English inside the Korean sentence ("Bohemian Rhapsody 가사 찾아줘"), user-composed free text becomes Korean. Within a row, every occurrence of the same original value uses one consistent form.

Korean purity (2026-10-04)

A deterministic pass (no LLM in the pass/fail decision) removed three defects a Korean training corpus should not carry. Originals are preserved; only the cleaned rows ship here.

defect	what / why	action
Chinese characters	the translator (Qwen3-30B-A3B) occasionally emitted Chinese (听起来, 归属되나요)	drop the row
instruction leak	translator meta-text left in the utterance (한국어로 번역된 사용자 메시지: …)	drop the row
register mixing	존댓말 and 반말 mixed in one speaker turn, judged by a Korean morphological analyzer (kiwi)	drop the row
emoji in tool arguments	emoji inside a function-call argument breaks consistency with the tool result	drop the row

Net: 193,149 → 181,894. Emoji and Markdown inside the natural-language text are kept — a previous variant that stripped emoji and flattened Markdown (for a text-to-speech use case) has been reverted; use that variant's pipeline step (tts_plain.py) if you need spoken-plain output.

Composition and upstream sources

file	rows	upstream dataset	upstream license
glaive.jsonl	42,661	glaiveai/glaive-function-calling-v2	Apache-2.0
xlam.jsonl	37,946	Salesforce/xlam-function-calling-60k	CC-BY-4.0
nemotron_tc.jsonl	33,063	nvidia/Nemotron-SFT-Agentic-v2 (tool_calling)	CC-BY-4.0
synth_apigen.jsonl	28,512	argilla/Synth-APIGen-v0.1	CC-BY-4.0
nemotron_ia.jsonl	19,416	nvidia/Nemotron-SFT-Agentic-v2 (interactive_agent)	CC-BY-4.0
toolace.jsonl	8,559	Team-ACE/ToolACE	Apache-2.0
robot.jsonl	5,667	none — generated in-house with a deterministic robot-environment simulator	CC-BY-4.0
apigen_mt.jsonl	4,588	Salesforce/APIGen-MT-5k	CC-BY-4.0
hermes.jsonl	1,482	NousResearch/hermes-function-calling-v1	Apache-2.0
total	181,894

Schema

field	type	description
id	str	same id as the English corpus — rows are joinable across the two
source	str	one of the 9 sources above
tools	str (JSON)	OpenAI-style function catalog, distractor-augmented, English
messages	str (JSON)	chat messages; assistant messages carry reasoning_content (Korean) alongside content and tool_calls
think_source	str	v24_react or v23_annotate — preserved from the English corpus

from datasets import load_dataset
ds = load_dataset("HBKenerzai/agentic-think-v2-react-hybrid-ko", split="train")

Method

• Translator: Qwen3-30B-A3B-Instruct-2507-FP8, temperature 0, one call per turn group (user message + that turn's assistant rounds, round count enforced).
• Five hard gates per turn (fail = drop the whole row); schema-enum / code-token argument values are protected from translation; a per-row post-pass unifies argument spellings across turns.
• Korean-purity pass applied afterward (see above).
• Quality audit (pre-purity): 360 rows stratified 40-per-source, judged side-by-side against the English original by a different model (Qwen3.8-27B) on five axes: 94.2% score ≥4 on all five.

Known limitations

• The translation gate had already dropped ~9% of the English corpus before the purity pass (concentrated in the longest multi-turn rows).
• Person names inside arguments may appear in Hangul when the Korean utterance says them that way (노아 ↔ Noah). Remap if your target API expects Latin-only fields.
• Register-mixing removal is morphology-rule-based (≈96% precision on a 180-row check); a few clean rows may have been dropped and a few mixes may remain.
• Reasoning is machine-translated from teacher-generated English reasoning; not human-verified.

Access

This dataset is gated: access is granted on request after manual review.

Provenance

Direct base: HBKenerzai/agentic-think-v2-react-hybrid (ENERZAi). Localization and purity pipeline: agentic-datagen data_gen/ko/, data_gen/purify_ko.py, data_gen/mix_rule.py (ENERZAi).