license: apache-2.0
task_categories:
- text-classification
- translation
language:
- en
- es
- de
- ja
tags:
- i18n
- localization
- translation-quality
- llm-as-judge
- benchmark
pretty_name: UIStringBench
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: public
path: data/public.jsonl
UIStringBench
A benchmark for detecting translation errors in software UI strings.
Machine and LLM translation of locale files fails in ways generic MT benchmarks don't measure: dropped or invented placeholders, collapsed plurals, and — hardest of all — fluent translations that state something the source doesn't (the Spanish that says "could not be saved" for "were saved successfully"). As of September 2026 there is no standardized public benchmark for this error class. UIStringBench is a first attempt at one.
What's in it
Each row is a source/translation pair from the software-UI domain with a gold label:
| field | description |
|---|---|
id |
stable pair id |
lang |
target language (es, de, ja) |
source |
English source string |
translation |
candidate translation |
label |
clean | mistranslation | omission | addition | glossary |
split |
public |
Public split: 228 pairs — 168 clean (drawn from real shipped production locale files, including
adversarial pairs whose content looks like instructions or JSON), 54 planted errors (18 per
category), and 6 glossary violations (deterministically checkable against the included
glossary.json; no model involved).
Held-out split: 60 pairs, private. Kept out of this repository as a contamination defense; leaderboard results report both splits separately. Composition mirrors the public split (42 clean / 6+6+6 planted, no glossary items).
Provenance — why you can trust the labels (and how far)
- The public corpus was committed to git (
Shipi18n/shipi18n@60d699b, 2026-08-16) before the judge that consumes it was written, and the evaluation thresholds were fixed before implementation. Pre-registration is the credibility spine of this dataset. - Labels were not revised after seeing model results. In published runs, roughly a quarter of
the judge's "false positives" turned out to be the judge being right about pairs mislabelled
clean— they are still counted against the judge, because a corpus you edit after seeing results is not a corpus. - Honest limitations: three language pairs; one domain (UI strings, mostly from one product's site); corpus and held-out split authored by the tool's maintainer with LLM assistance. Planted errors are deliberate and unambiguous — real-world drift is subtler. Treat scores as an engineering QA signal, not a validated academic metric (the GEMBA authors' caution applies here too).
Related work
- GEMBA-MQM (Kocmi & Federmann, WMT 2023) and successors (RUBRIC-MQM, GEMBA v2): LLM-as-judge MT evaluation on generic sentence pairs. UIStringBench differs in domain (locale files: placeholders, plurals, short ambiguous strings) and in shipping as a runnable dev-tool harness.
- Last Translation Benchmark: live, crowd-sourced hard-to-translate examples — general MT, not UI strings or structural failure modes. Adjacent, not overlapping.
- COMET/xCOMET: learned MT metrics on sentence pairs; multi-GB downloads, no locale-file awareness.
Running it
The harness lives in the Shipi18n repo under
evals/semantic/:
git clone https://github.com/Shipi18n/shipi18n && cd shipi18n
JUDGE_MODEL=<your-model> node evals/semantic/run.mjs
The judge design being measured: 3-pass majority vote, MQM-style category output (mistranslation / omission / addition) rather than scalar scores, structural-first filtering. Item-level results are published with each leaderboard release.
Versioning
v1 (2026-09). New languages and formats (gettext, ICU plurals) will widen future versions; saturated versions will be re-cut rather than silently edited. Maintainer: the Shipi18n project. Issues and label disputes: GitHub issues.