한국어 Contact us

Home Services RLHF and post-training data

Korean RLHF and post-training data

We write and review Korean post-training data for global AI developers and AI data platforms: prompts, reference answers, preference rankings, rewrites, tool-calling dialogues and instructions with constraints.

Contact us See examples

What you receive

  • 01

    Data in your schema

    Prompts, answers, rankings, rewrites or dialogues in JSONL, CSV or spreadsheets with your field names, or entered directly in your tool.

  • 02

    A reason for each judgment

    Every ranking, rating and rewrite comes with a short note naming the criterion behind it, in English or Korean.

  • 03

    A Korean rulings list

    Decisions on politeness level, spacing, number formats and loanwords, written down once and applied across the whole batch.

  • 04

    An issue report

    Unclear guideline points, invalid tasks and conflicting criteria we found, with the items affected and how we ruled.

How a project runs

  1. 1

    Send the brief

    Email us the model or product, the data type, and a rough volume and timeline. Attach any guidelines you have.

  2. 2

    Agree scope and format

    We read your guidelines, send questions on unclear points, then agree scope, file format and schedule with you.

  3. 3

    Pilot batch

    We deliver a small sample first. You review it, and we settle open rulings before full volume.

  4. 4

    Write and review

    We write and review the full set to one standard, then check format and fields with our own tools.

  5. 5

    Delivery with notes

    You receive the data, a note per judgment and an issue report. Items you send back are fixed within the agreed scope.

What we check

Prompts and tasks

  • Sounds like a real Korean user, not a translation
  • Every constraint can be checked from the answer
  • Constraints never contradict each other
  • Adversarial tasks expose a real model failure
  • No real personal data or copied text

Reference answers and rewrites

  • Meets every instruction and constraint in the prompt
  • Facts, numbers and dates checked against a source
  • Answers the question first, without padding
  • No needless refusals and no unsafe content
  • Tool calls: right tool, valid arguments, no invented results
  • Rewrites keep what was right and fix only the faults

Preference rankings

  • Accuracy, instruction and safety failures outrank style
  • Length alone never decides the order
  • Ties only when the differences are trivial or cancel out
  • Each ranking names the criterion that decided it
  • Same issue, same ruling across the whole batch

Korean writing

  • One politeness level from start to finish
  • Register fits the reader: formal, polite or casual
  • No translationese, such as wordy passive phrasing
  • Particles fit the word before them (을/를, 이/가)
  • Standard spacing and loanword spelling
  • Korean date, time and number formats

Worked examples

Synthetic data, made for this page.

Example

Ranking two summaries, then rewriting the winner

A user asks for a three-line summary of meeting notes, and we rank two model answers.

  1. Prompt

    회의록을 세 줄로 요약해 주세요.- 출시일: 인증 서류 보완 때문에 11월 20일로 연기- 디자인 시안: 금요일까지 수정- 가격: 다음 회의에서 결정- 다음 회의: 11월 3일 오전 10시

    Summarize the meeting notes in three lines. Launch: moved to 20 Nov because the certification papers need fixing. Design draft: revise by Friday. Price: decide at the next meeting. Next meeting: 3 Nov, 10 a.m.

  2. Answer A

    1. 인증 서류 보완 때문에 출시일이 11월 20일로 연기되었습니다.2. 디자인 시안에 대한 수정 작업이 금요일까지 진행될 예정입니다.3. 가격은 11월 3일 오전 10시 회의에서 결정될 예정입니다.

    1. The launch has been postponed to 20 Nov because the certification papers need fixing. 2. Revision work regarding the design draft is scheduled to be carried out by Friday. 3. The price is scheduled to be decided at the 3 Nov, 10 a.m. meeting.

  3. Answer B

    1. 출시일이 11월 20일로 연기되었습니다.2. 연기 이유는 인증 서류 보완입니다.3. 디자인 시안은 금요일까지 수정합니다.4. 가격은 11월 3일 오전 10시 회의에서 정합니다.5. 일정이 빠듯하니 디자인팀과 미리 상의해 두시면 좋겠습니다.

    1. The launch has been postponed to 20 Nov. 2. The reason is that the certification papers need fixing. 3. The design draft will be revised by Friday. 4. The price will be set at the 3 Nov, 10 a.m. meeting. 5. The schedule is tight, so it would be good to talk with the design team early.

  4. Rewrite of A

    1. 인증 서류 보완 때문에 출시일이 11월 20일로 연기되었습니다.2. 디자인 시안은 금요일까지 수정합니다.3. 가격은 11월 3일 오전 10시 회의에서 정합니다.

    1. The launch has been postponed to 20 Nov because the certification papers need fixing. 2. The design draft will be revised by Friday. 3. The price will be set at the 3 Nov, 10 a.m. meeting.

VerdictA preferred, then rewritten

Instruction following outranks style: B reads well but runs to five lines and adds advice nobody asked for. A keeps every fact in three lines, so it wins, and the rewrite only fixes its translated, wordy phrasing in lines 2 and 3.

Example

Reference answer, checked constraint by constraint

We wrote a travel prompt with three checkable constraints, then checked a draft answer against each one.

  1. Prompt

    부모님과 갈 제주도 2박 3일 일정을 짜 주세요. 부모님이 보실 거라 알파벳은 쓰지 말고, 하루에 두 곳까지만 넣어 주세요. 끝에 저한테 질문은 하나만 해 주세요.

    Plan a three-day, two-night Jeju trip for me and my parents. They will read it, so use no Latin letters, and include at most two places a day. Ask me only one question at the end.

  2. Draft answer

    1일차: 동문시장, 사라봉2일차: 성산일출봉, 광치기 해변, 섭지코지3일차: 협재 해변, 한림공원숙소는 BnB가 좋을까요?

    Day 1: Dongmun Market, Sarabong. Day 2: Seongsan Ilchulbong, Gwangchigi Beach, Seopjikoji. Day 3: Hyeopjae Beach, Hallim Park. Would a BnB be good for lodging?

  3. Constraint check

    Two places a day: fail, day 2 has three.No Latin letters: fail, 'BnB'.One question at the end: pass.

  4. Reference answer

    1일차: 동문시장, 사라봉2일차: 성산일출봉, 섭지코지3일차: 협재 해변, 한림공원숙소는 민박과 호텔 중 어디가 좋으세요?

    Day 1: Dongmun Market, Sarabong. Day 2: Seongsan Ilchulbong, Seopjikoji. Day 3: Hyeopjae Beach, Hallim Park. For lodging, would you prefer a minbak (family-run guesthouse) or a hotel?

VerdictDraft fails two constraints

Each constraint is checked against the text alone: day 2 lists three places, and 'BnB' is written in Latin letters. The reference answer keeps the plan and fixes only those two points.

Data and formats

  • Single and multi-turn chats
  • Prompt and reference answer pairs
  • Pairwise and ranked preferences
  • Rewrites paired with the original
  • Tool-calling dialogues with JSON schemas
  • Math Q&A with worked solutions
  • JSONL, JSON, CSV and spreadsheets
  • Your annotation tool or platform

Questions

Do you write the data, or only review it?

Both. We write prompts, reference answers, rewrites and tool-calling dialogues, and we review data from other writers or from models against the same checks.

Do you follow our guidelines or your own?

Yours come first. The checks on this page are our defaults for gaps your guidelines leave open. Without guidelines, we propose short Korean rules, such as politeness level and number formats, and start once you approve them.

Can tool-calling dialogues use our own tool schemas?

Yes. Send your tool definitions, and we write dialogues that call them with valid arguments, including several calls in one turn and calls that use an earlier result. Every call is checked against your schema before delivery.

Can you write prompts that current models get wrong?

Yes. We write adversarial tasks built on stacked constraints, relative dates such as 'this Saturday', or switches in politeness. When we can run the model you target, a task counts only if the model actually fails it.

Do you also review math or chemistry content?

Yes. We review solutions and final answers in math Q&A, and check identifiers and structural formulas in chemistry data against the literature.

Other services

LLM and agent evaluation Red teaming and safety evaluation Localization QA In-house QA tools

Contact

Tell us about your project

Write in English or Korean, and include these four points if you can.

  1. 1Model, agent or product
  2. 2Type of work
  3. 3Rough volume and timeline
  4. 4Existing guidelines, if any
contact@marudata.io

The address opens a draft in your mail app. We then agree scope and schedule with you.