README.md

---
task_categories:
- translation
language:
- en
- ko
tags:
- game
- text
- translation
pretty_name: KUEST
size_categories:
- 1K<n<10K
---

# **KUEST**: **K**orean game **U**sage **E**valuation **S**uite for llm **T**ranslation

A WMT26 MT test suite focusing on translation blind spots that surface when translating English into Korean in the game domain.

Translation direction: English → Korean (`en` → `ko_KR`)

## Participant information

- **Test suite name:** **KUEST** (**K**orean game **U**sage **E**valuation **S**uite for llm **T**ranslation)
- **Institution:** Pseudo Lab
- **Contact:** Minjoo Choi — mjchoi0831@gmail.com

## Team Members
- **Minjoo Choi**
- Jaejun Jung
- Gunwoo Han
- Junhyeop Kim
- Yiji Yoon
- Hyemi Kim
- Yeonseo Cho
- Jinwoo Jang

## Overview

- **Language direction:** English → Korean (`src_lang` = `en`, `tgt_lang` = `ko_KR`)
- **Domain:** Video games (Cataclysm: Dark Days Ahead 2025 `ko_25.po`, Battle for Wesnoth PO modules)
- **Submission file:** `KUEST.jsonl` (JSON Lines, UTF-8)
- **Items:** 1,880
- **Source tokens:** ~20,842 (whitespace), well under the 1M per-participant limit
- **License:** Submitted source sentences may be made public for research purposes.

## What it probes

KUEST targets phenomena where fluent MT output can still be wrong for Korean in the game domain. It groups items into four categories:

| Category | Items | Focus |
|---|---|---|
| **TA** — Terminology & Accuracy | 331 | Glossary canonical compliance; polysemy disambiguation by context |
| **PP** — Persona & Pragmatics | 472 | Speech-register/honorific consistency by speaker persona; **dual-prompt** (baseline vs. persona-enriched regime R) |
| **TR** — Translation Robustness | 507 | Placeholder reordering, tag preservation, length constraints |
| **CT** — Cultural Transcreation | 570 | Honorifics, onomatopoeia, idioms, naming, hate speech, etc. |

Items are ordered TA → PP → TR → CT.

### PP dual-prompt (regime R)

PP now ships the same 236 speaker utterances under **two prompt regimes** so the effect of persona context on the translation can be measured:

- **baseline** — the fixed prompt `Translate from en to ko_KR.` (subcategories `PP-persona_pragmatics`, `PP-ambiguous`; 236 rows).
- **regime R** — an English prompt enriched with speaker persona/relationship facts (speaker name, role tier, in-game unit type, speech act, and social-standing relationship), without prescribing Korean endings (subcategories `PP-persona_pragmatics-regime_r-001-a … -097-b` and `PP-ambiguous-regime_r-001-a … -021-b`; 236 rows).

## Record format

Every line contains exactly the seven WMT fields:

```json
{"dataset_id": "KUEST", "doc_id": "en-ko_KR_#_KUEST_#_CT-idiom_proverb-001", "src_lang": "en", "tgt_lang": "ko_KR", "prompt": "Translate from en to ko_KR.", "src_text": "Your skill in speaking to other people. Covers ability in boasting, flattery, threats, persuasion, lies, and other facets of interpersonal communication. Works best in conjunction with a high level of intelligence.", "ref_text": "사람들과 대화하는 기술입니다. 자랑, 아첨, 위협, 설득, 거짓말 등 사람들과의 소통 전반에 영향을 미칩니다. 지능이 높은 캐릭터와 결합하면 금상첨화일 것입니다."}
```

- `doc_id` pattern: `en-ko_KR_#_KUEST_#_<category>-<subcategory>-<NNN>`; numbering restarts at `001` per subcategory (3-digit zero-pad). The PP category splits each speaker pair into two rows with `-a` / `-b` suffixes; the regime R rows reuse the source subcategory stem with a `-regime_r-` infix (`PP-persona_pragmatics-regime_r-…`, `PP-ambiguous-regime_r-…`), keeping the source split and the original pair numbering.
- `prompt` is fixed (`Translate from en to ko_KR.`) for every row **except** the 236 regime R rows (the `*-regime_r-*` subcategories), which carry persona-enriched prompts (see PP dual-prompt above).
- `ref_text` holds the Korean reference (never null in this suite); it is identical across both PP regimes.

## Text normalization

- Paragraph/line breaks are encoded as `\n\n` (per the WMT format note).
- Tabs and other control characters are replaced with a single space.
- Placeholders and tags (`%s`, `{n}`, `<span ...>`, `$player_…`) are preserved; source↔reference placeholder parity was verified for all TR items.


# KUEST — Dataset Licensing

KUEST is distributed as a single data file, `KUEST.jsonl`. Its items are derived
from open-source game-localization data released under **two different, mutually
incompatible copyleft licenses** (GPL-2.0 and CC BY-SA 3.0). `KUEST.jsonl` is
therefore **not under a single license**: it is a *collection* in which **each
record is independently licensed** according to its `license` field.

> Even though the data itself is one file, a compliant release must also ship this
> notice and copies of the two license texts — see "Redistribution & compliance".

## Component licenses

| Upstream source | Records | License (SPDX) | Upstream repository |
|---|---:|---|---|
| CDDA(Cataclysm-DDA) | 855 | `CC-BY-SA-3.0` | https://github.com/CleverRaven/Cataclysm-DDA |
| Wesnoth(Battle for Wesnoth) | 941 | `GPL-2.0` | https://github.com/wesnoth/wesnoth |
| **Total** | **1796** | *(mixed collection)* | |

Each record in `KUEST.jsonl` carries:

- `source` — the upstream project the record's text is derived from.
- `license` — the SPDX identifier governing that record's `src_text` and `ref_text`.

Filter by the `license` field to obtain a single-license subset.

## Attribution

**Cataclysm-DDA** — Source and reference-translation text for 855 records is derived
from *Cataclysm: Dark Days Ahead*, © the Cataclysm-DDA contributors, licensed under
the Creative Commons Attribution-ShareAlike 3.0 Unported License (CC BY-SA 3.0).
License: https://creativecommons.org/licenses/by-sa/3.0/

**Battle for Wesnoth** — Source and reference-translation text for 941 records is
derived from *The Battle for Wesnoth*, © the Battle for Wesnoth Project, licensed
under the GNU General Public License, version 2 or later (GPL-2.0-or-later).
License: https://www.gnu.org/licenses/old-licenses/gpl-2.0.html

## Compilation

The selection, structuring, and original fields authored by the KUEST authors
(e.g., the `doc_id` scheme and `prompt` field) are offered under **`CC-BY-4.0`**.
This applies **only** to the KUEST authors' original material and does **not**
relicense the `src_text`/`ref_text`, which remain under their component licenses
above. Where upstream text was modified (e.g., the gender-neutral reference
corrections), the modified text is a *derivative* of the copyleft source and stays
under that source's license (`CC-BY-SA-3.0`).

## Redistribution & compliance

If you redistribute `KUEST.jsonl` or a derivative:

- **Keep each record under its component license** (the `license` field). Do not
  relicense the whole file under a single or permissive license.
- **Ship the license texts.** Include this notice and verbatim copies of the
  CC BY-SA 3.0 and GPL-2.0 license texts alongside the data (e.g., in `licenses/`).
  Both licenses require the license to accompany the work; GPL requires including a
  copy of its text.
- **Preserve attribution** to both upstream projects.
- **ShareAlike / copyleft.** Distribute any adaptation of a record under the same
  (or a compatible) license as that record.
- **GPL source form.** For the Battle for Wesnoth records, the text itself is the
  corresponding source form, provided under GPL-2.0-or-later.

**More conservative option.** The single mixed file is a *labeled collection* and is
defensible, but if you want the airtight "separable aggregation" posture, split the
data by license into two files — e.g., `KUEST-cdda.jsonl` (`CC-BY-SA-3.0`) and
`KUEST-wesnoth.jsonl` (`GPL-2.0-or-later`) — instead of one mixed file. This removes
any ambiguity about whether the file is a single combined work.

## Note

This is a *mere-aggregation* collection: records under different licenses are
bundled together, but each remains independently licensed via its `license` field.
This document is **not legal advice** — verify license compatibility and your
redistribution obligations with your institution before public release.