Sky-Kim/convfill-dataset-ko

ConvFill Dataset, Korean

This is a Korean version of the ConvFill dataset. All 8,443 conversations of the original (290,571 aligned phrase pairs) were translated into Korean and checked automatically so that their structure is unchanged: the same turns, the same number of thoughts and response items per turn, and <sil> in exactly the same positions. The files, field names and line order match the original, so the original training code reads them as they are.

ConvFill comes from the paper Thinking While Speaking: Inference-Time Knowledge Transfer for Responsive and Intelligent Conversational Voice Agents (Srinivas, Englhardt, Iyer and Patel). It trains a small on-device Talker model to keep talking while a large cloud Reasoner model is still working. The Reasoner sends a stream of short knowledge chunks, or <sil> when nothing has arrived yet. For each item, the Talker says one phrase: it turns a knowledge chunk into natural speech, and it covers a <sil> with filler that does not invent facts.

Contents

File	Domain	Conversations	Turns	Phrase pairs	<sil> share
data/conversations_advice.jsonl	Advice	1,000	8,308	36,710	34%
data/conversations_assistant.jsonl	General assistant	1,000	8,408	37,495	34%
data/conversations_dstc8_scaffold.jsonl	Task-oriented, localized (DSTC8 / SGD)	2,438	24,184	68,617	54%
data/conversations_education.jsonl	Education	1,000	8,421	37,153	34%
data/conversations_medical.jsonl	Medicine	1,000	8,400	37,246	34%
data/conversations_planning.jsonl	Planning	1,005	8,474	37,488	34%
data/conversations_support.jsonl	Customer service	1,000	8,313	35,862	34%
Total		8,443	74,508	290,571	38%

data/conversations_dstc8_scaffold.jsonl keeps the top-level scaffold_metadata (SGD service and dialogue id) from the original. The other files do not have that field, as in the original; when loaded with datasets it is None for them. Every line also has source_line, the 0-based line number of the same conversation in the original file.

Format

One conversation per line. Each turn has a user utterance and two lists of equal length: thoughts[i] is what the Reasoner had produced at step i, and response[i] is what the Talker says at that step. The first turn of one conversation, pretty-printed:

{
  "conversation": [
    {
      "user": "삼월 오일 날씨는 어떨까요?",
      "thoughts": [
        "<sil>",
        "어느 지역의 날씨를 찾아볼까요?"
      ],
      "response": [
        "삼월 오일 날씨를 확인해 드릴게요.",
        "어느 지역의 날씨를 찾아볼까요?"
      ]
    }
  ],
  "scaffold_metadata": {
    "service": "Weather_1",
    "dialogue_id": "91_00011"
  },
  "source_line": 1
}

Usage

from datasets import load_dataset

ds = load_dataset("Sky-Kim/convfill-dataset-ko", split="train")
print(ds[0]["conversation"][0]["response"])

To train with the original code, download data/ and run its preprocessing step on that folder, the same way as for the English data:

python3 dataset_gen/dataset_preprocess.py --user_tag user --thoughts_tag thoughts \
  --responder_tag response --base_data_dir data --output_path convfill_ko.jsonl --include_history

How it was made

• Translation guidelines:
  • Keep every list the same length and <sil> at the same positions. response[i] must carry the same facts as thoughts[i].
  • user reads like a speech-to-text transcript. response is spoken polite Korean (해요체) for a voice assistant: no markdown, emoji, parentheses or URLs, since it goes to a TTS engine. thoughts are short, plain information sentences.
  • All numbers, times, dates, prices and phone numbers are written out in Hangul, the way they are read aloud ("오후 세 시 반", "만 오천 원"), in both thoughts and response. This covers measurements, ratings and decimals too ("일 점 삼 센티미터"). It matches a Reasoner prompted to spell out numbers, and avoids TTS number-reading errors.
  • Titles of books, songs, albums and films use the official Korean title, or are written in Hangul ("렌카의 블루 스카이즈"). Latin script is kept only for code, commands, file names, acronyms, and English that is itself the subject, such as UI labels or example sentences in an English lesson.
  • The DSTC8/SGD conversations are localized to Korea: Korean cities and districts, Korean business names and addresses, Korean phone numbers, prices in won, and Korean sports leagues and artists. A localized name stays identical everywhere it appears in a conversation. The other domains are translated faithfully. Medical, legal and safety facts are kept as stated.
• Automatic checks: every translated conversation had to pass all of these:
  • the same turn count, list lengths and <sil> positions as the source
  • no empty strings, and no stray <sil> inside text
  • no leftover English prose (Latin script stays only where the original has it on purpose, such as code, commands, UI labels and English example sentences in a grammar lesson)
  • no numbers left in digits in thoughts or response, outside code and product names such as GA4
  • no markdown or other non-spoken characters in response
• Retries: a conversation that failed a check was translated again on its own until it passed. 233 of 8,443 conversations (2.8%) failed on the first try, mostly because the thoughts and response lists came back with different lengths. Every conversation of the original is included; none was dropped.
• Pilot: 50 conversations were translated first and read by hand before the rest. This led to two rule changes: a reminder for day numbers (one "11th" had become "십일" instead of "십일일"), and localizing whole entities at once (one US football match had kept its US team names but moved to Seoul).

Limitations

• Localized entities are fictional: the Korean businesses, addresses, phone numbers and prices in the DSTC8 part were invented during localization. They are consistent within each conversation but do not refer to real places.
• Spelled-out numbers: numbers appear in Hangul, not digits. A model trained on this data will tend to speak numbers that way, and expects the Reasoner to send them that way too.
• Medical and other expert content: this was translated, not reviewed by clinicians or other experts. It is training data for conversational style, not a source of advice.

License and attribution

This dataset is released under the Creative Commons Attribution-ShareAlike 4.0 International license (CC BY-SA 4.0), the same license as the original. If you share or adapt it, give appropriate credit, indicate what you changed, and release your adaptation under the same license.

• Original dataset: this is a translation, and therefore an adaptation, of the ConvFill dataset by Vidya Srinivas, Zachary Englhardt, Vikram Iyer and Shwetak Patel (University of Washington), CC BY-SA 4.0.
• Changes made: all text was translated from English to Korean, numbers were written out in Hangul, the DSTC8 conversations were localized to Korea as described above, and a source_line field was added.
• Upstream data: the DSTC8 part of the original is built on the Schema-Guided Dialogue dataset by Google Research, CC BY-SA 4.0.

If you use this dataset, please cite the original work:

@misc{srinivas2026thinkingspeakinginferencetimeknowledge,
      title={Thinking While Speaking: Inference-Time Knowledge Transfer for Responsive and Intelligent Conversational Voice Agents},
      author={Vidya Srinivas and Zachary Englhardt and Vikram Iyer and Shwetak Patel},
      year={2026},
      eprint={2511.07397},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.07397},
}