Maru Data · 마루데이터 Korean-language evaluation and post-training data for LLMs and AI agents.
Evaluation, red teaming, RLHF data and localization QA in Korean, for global AI developers and for AI data and evaluation platforms.
Fig. 1 Example ruling
Illustrative · 예시UserCasual · 반말
내일 오후에 회의 하나 잡아 줘.
“Book a meeting for tomorrow afternoon.”
AgentFormal, then casual
내일 오후로 회의를 잡아 두었습니다. 시간이 바뀌면 알려 줘.
“The meeting is booked for tomorrow afternoon. Let me know if the time changes.”
- Ruling
- Answering a casual user in formal speech is acceptable. The drop to casual speech in the second sentence is an error.
- Correct
- 시간이 바뀌면 알려 주십시오.
- Applies to
- Every item in the batch with the same pattern. If the guideline does not cover the case, it is raised with the project.
-
Korean-locale LLM and AI agent evaluation
Rubric-based review of model answers and agent trajectories in Korean: correctness, instruction following, and whether the Korean reads naturally. Agent runs are judged step by step, as well as on the final answer.
-
RLHF and post-training data in Korean
Korean prompts, reference answers, preference rankings and rewrites for fine-tuning and RLHF, written to the project’s guidelines and kept consistent across the batch.
-
Red teaming and safety evaluation in Korean
Adversarial testing of LLMs in Korean, including prompt injection and attacks that depend on Korean phrasing, slang or mixed Hangul and Latin text. Findings are reported with the prompts that produced them.
-
Localization QA
Review of Korean UI text, documentation and model output for accuracy, terminology, tone and local conventions such as honorifics, dates and number formats.
-
Python tooling for evaluation pipelines
Python tools for parsing agent trajectories, reconciling steps and auditing a batch before it is submitted.
A score only means something if the same ruling holds from the first item in a batch to the last.
-
Consistent rulings across the batch
Calibration is part of the work. Once an edge case is decided, the same ruling is applied to every comparable item in the batch.
-
Recurring problems are reported
When the same problem keeps coming up, such as an unclear guideline, a broken task or a rubric that contradicts itself, it is reported to the project as a systemic issue so it can be fixed at the source.
-
Checked with in-house tooling
Agent work runs through Python tools written for it. Trajectories are parsed into steps, the steps are reconciled, and the batch is audited before submission.
Fig. 2 In-house tooling
Schematicbatchparsereconcileauditsubmission
- parse
- Read each agent trajectory into ordered steps.
- reconcile
- Check that step-level judgments line up with the trajectory.
- audit
- Check the whole batch for gaps and inconsistencies before it goes out.
회사 소개
About
Korean-language evaluation work since 2024, registered as Maru Data in 2026.
Maru Data is run by its founder, Hongmin Lim. Since 2024 Hongmin has worked as an independent contractor on Korean-language LLM evaluation, RLHF and post-training data, and red-teaming projects for global AI developers, mostly through international AI data and evaluation platforms.
On 13 October 2026 the business was registered as a sole proprietorship in the Republic of Korea. It is based in Incheon.
- Company
- Maru Data (마루데이터)
- Legal form
- Sole proprietorship, Republic of Korea
- Founder
- Hongmin Lim
- Registered
- 2026-10-13
- Location
- Incheon, Republic of Korea
- Independent work
- Since 2024
- contact@marudata.io
- Website
- marudata.io
Confidentiality
Client names are confidential. This site describes past work by type and does not name clients or the platforms involved.
문의
Contact
Write in English or Korean. A useful first email covers:
- the model, agent or product involved
- the type of work: evaluation, post-training data, red teaming, localization QA or tooling
- rough volume and timeline
- any guidelines or rubrics that already exist
Incheon, Republic of Korea · Korea Standard Time (KST)