Datasets:
metadata
pretty_name: zh_translation_benchmark
language:
- zh
- en
language_bcp47:
- zh-TW
- cmn-TW
- en-US
- zh-CN
- zh-Hant-HK
license: cc-by-4.0
task_categories:
- translation
size_categories:
- 1K<n<10K
tags:
- taiwan
- traditional-chinese
- zh-tw
- taiwan-mandarin
- localization
- machine-translation
- mt-benchmark
- translation-evaluation
- code-mixed
- long-document
- synthetic-data
configs:
- config_name: full
data_files: data/full.jsonl
zh_translation_benchmark
zh_translation_benchmark is a 2,000-example synthetic benchmark for evaluating whether a translation or rewriting model can produce natural Taiwan Traditional Chinese (zh-TW) from English, Mainland Chinese, Hong Kong Traditional Chinese, Cantonese-style written Chinese, or code-mixed English/Chinese documents.
The dataset now exposes a single Hugging Face subset/config: full. It is not split into dev and test files. The full 2,000 rows are loaded together from data/full.jsonl. The original row-level split column is kept as metadata for downstream analysis, but the Hub dataset itself is published as one complete subset.
Credits
This dataset is created and released by OpenFormosa.
Loading
from datasets import load_dataset
dataset = load_dataset("OpenFormosa/zh_translation_benchmark", "full")
# Hugging Face stores a single data file under the conventional "train" split.
full = dataset["train"]
print(full.num_rows) # 2000
print(full[0])
If you want a plain Dataset object:
from datasets import load_dataset
full = load_dataset("OpenFormosa/zh_translation_benchmark", "full", split="train")
Version Summary
This v2 release merges the earlier v0 sentence-level localization set with a v1 long-document and code-mixed localization set.
- v0: sentence-level locale-sensitive translation from English, zh-CN, and zh-Hant-HK into Taiwan Traditional Chinese.
- v1: paragraph and long-document localization with code-mixing, Markdown, placeholders, cross-paragraph terminology consistency, and broader domain coverage.
- v2: unified schema with one official
fullsubset containing all 2,000 rows.
Dataset Summary
| Property | Value |
|---|---|
| Rows | 2,000 |
| Unique dedup hashes | 2,000 |
| Domains | 56 |
| Domain groups | 18 |
| Source versions | v0: 1,000; v1: 1,000 |
| Published subset/config | full |
| Published split files | none; one full JSONL file |
Direction Distribution
| Direction | Count |
|---|---|
| English -> Taiwan Traditional Chinese | 950 |
| zh-CN -> Taiwan Traditional Chinese | 550 |
| mixed English/Chinese -> Taiwan Traditional Chinese | 300 |
| zh-Hant-HK -> Taiwan Traditional Chinese | 200 |
Length Distribution
| Length bucket | Count |
|---|---|
| Sentence | 1,500 |
| Paragraph | 300 |
| Long document | 200 |
What This Dataset Tests
Many systems can output Traditional Chinese characters but still leave Mainland Chinese, Hong Kong wording, or translated English workplace jargon in the sentence. This benchmark targets those errors directly.
Examples of intended localization behavior:
| Non-Taiwan or source term | Preferred Taiwan-localized handling |
|---|---|
用户 |
使用者 |
默认 |
預設 |
缓存 |
快取 |
服务器 |
伺服器 |
信息 |
資訊 |
数据 |
資料 |
地铁 |
捷運 in Taiwan contexts |
公交 |
公車 |
出租车 / 的士 |
計程車 |
外卖 |
外送 |
身份证 / 身份證 |
身分證 |
医保卡 |
健保卡 |
港鐵 in Taiwan-localized transport copy |
捷運 |
八達通 in Taiwan-localized app copy |
悠遊卡 |
v2 also tests:
- Cross-paragraph terminology consistency
- Markdown and placeholder preservation
- Code-mixed workplace text
- UI and product copy localization
- Public-service, medical, legal, finance, and customer-support register
- Whether English terms should be retained or translated in Taiwan contexts
Domains
The full set covers 56 domains across 18 domain groups. Major domains include:
| Domain | Count |
|---|---|
software_it |
130 |
government_public_service |
72 |
workplace_business |
67 |
education |
66 |
ecommerce_product_ui |
61 |
zhcn_software_it |
60 |
zhcn_ecommerce |
50 |
zhcn_workplace |
50 |
zhcn_public_health_gov |
50 |
zhcn_daily_transit_food |
50 |
zhcn_education_media |
40 |
product_ui |
37 |
cybersecurity |
37 |
ecommerce |
37 |
fintech_banking |
37 |
insurance |
37 |
healthcare_clinic |
37 |
healthcare |
36 |
marketing_social_media |
35 |
medical_device |
35 |
legal_privacy |
35 |
academic_research |
35 |
workplace_hr |
35 |
Use the domain, domain_group, domain_zh, and subdomain columns for more detailed breakdowns.
Files
| Path | Description |
|---|---|
data/full.jsonl |
Single 2,000-row JSONL file used by the Hugging Face loader |
data/full.csv |
Single 2,000-row CSV file |
full/tw_localization_benchmark_v2_2000.csv |
Complete 2,000-row CSV |
full/tw_localization_benchmark_v2_2000.jsonl |
Complete 2,000-row JSONL |
full/tw_localization_benchmark_v2.xlsx |
Workbook with Overview, Examples_2000, Core_1000, Rubric, Locale_Policy, Glossary, Eval_Prompts, Auto_Checks, Model_Slots, and ChangeLog sheets |
full/tw_localization_benchmark_v2_manifest.json |
Manifest with counts, checksums, schema, and historical core selection rule |
full/tw_localization_benchmark_v2_README.md |
Original v2 local package README |
Schema
Each example contains:
| Field | Description |
|---|---|
id |
Stable v2 ID, such as TWLOC-V2-0001 |
legacy_id |
Original v0 or v1 ID |
version_source |
Source version: v0 or v1 |
core_1000 |
Whether the row belonged to the historical core set |
split |
Original dev/test marker retained as row metadata |
source_lang |
Normalized source language bucket |
source_locale_raw |
Original source locale label |
target_lang |
Target language, zh-TW |
direction |
Translation or rewriting direction |
domain_group |
High-level domain group |
domain |
Domain label |
domain_zh |
Chinese domain label |
subdomain |
Optional subdomain |
genre |
Genre or text style |
doc_type |
Document type |
length_bucket |
sentence, paragraph, or long_doc |
source_len_chars |
Source length in characters |
target_len_chars |
Reference length in characters |
code_mix_level |
Code-mixing level |
code_mix_ratio_est |
Estimated code-mix ratio |
code_mix_pattern |
Code-mixing pattern |
context_dependency |
Required cross-segment context, if any |
phenomenon |
Main evaluated phenomenon |
phenomena_tags |
Additional phenomenon tags |
difficulty |
Estimated difficulty from 1 to 5 |
stress_axes |
Main stress-test axis |
source_text |
Input text |
reference_tw_zh |
Reference Taiwan Traditional Chinese output |
acceptable_variants |
Optional acceptable alternate wording |
must_preserve |
Terms, placeholders, URLs, Markdown, numbers, names, or formatting that must be preserved |
avoid_terms |
Row-level Mainland, Hong Kong, or unnatural terms to avoid |
locale_policy |
Taiwan localization policy for the row |
eval_notes |
Evaluation notes |
comment |
Optional comment |
data_origin |
Data provenance marker |
review_status |
Current review status |
dedup_hash |
Deduplication hash |
release_tag |
Release tag |
created_date |
Creation date |
Example
{
"id": "TWLOC-V2-0001",
"legacy_id": "TWLOC-0001",
"version_source": "v0",
"core_1000": "TRUE",
"split": "dev",
"source_lang": "en",
"source_locale_raw": "en-US",
"target_lang": "zh-TW",
"direction": "en→zh-TW",
"domain_group": "technology",
"domain": "software_it",
"domain_zh": "軟體與資訊技術",
"length_bucket": "sentence",
"source_text": "Users reported that the admin console keeps spinning after the latest release.",
"reference_tw_zh": "使用者回報，最新版本釋出後，後台管理介面一直卡在載入中。",
"avoid_terms": "用户、发布、一直转圈",
"release_tag": "TW-Local-MT-Bench-v2"
}
Recommended Inference Prompt
請將以下文字或文件翻譯或轉寫成台灣本土慣用的繁體中文。請保留原文的事實、專有名詞、數字、日期、貨幣、Markdown、HTML tag、placeholder、URL、email、SLA、KPI、API 名稱與格式。若原文中英文夾雜，請判斷哪些英文是產品名、程式介面、檔名、欄位名或常見縮寫而應保留，哪些只是英文工作口語而應翻成自然台灣中文。請保持全文術語一致，避免中國大陸或香港用語。只輸出譯文。
來源文字：{source_text}
Evaluation
Suggested 100-point human evaluation rubric:
| Dimension | Weight | What to check |
|---|---|---|
| Adequacy | 30 | Source facts, conditions, negation, numbers, entities, and causality |
| Taiwan localization | 25 | Taiwan vocabulary, institutions, tone, punctuation, and document style |
| Terminology consistency | 15 | Consistent terminology across paragraphs and documents |
| Code-mix handling | 10 | Correctly keeping product/API/file/field names while translating casual English into natural Taiwan Chinese |
| Fluency and register | 10 | Natural, readable output matching customer support, legal, medical, UI, government, or casual register |
| Format and constraint preservation | 10 | Markdown, placeholders, URLs, HTML, numbers, currencies, tables, and bullets |
LLM-Assisted Judging Prompt
你是台灣繁體中文翻譯評審。請比較來源文字、參考譯文、模型輸出。請輸出 JSON：adequacy_0_30、tw_localization_0_25、terminology_consistency_0_15、code_mix_handling_0_10、fluency_register_0_10、format_constraints_0_10、total_0_100、major_errors、minor_errors、error_spans、comment。請優先懲罰語意錯誤、簡體字、中國大陸或香港用語、台灣制度名詞錯誤、術語前後不一致、Markdown/placeholder/URL/數字或專有名詞改錯。
Intended Uses
This dataset is suitable for:
- Taiwan-localized machine translation evaluation
- Prompt testing for
zh-TWtranslation systems - Comparing open-weight or open-source LLMs on Taiwan localization
- Measuring Mainland/Hong Kong terminology leakage in Traditional Chinese output
- Testing code-mixed and long-document localization
- Building human annotation workflows and internal model regression tests
- Stress-testing Markdown, placeholder, URL, and terminology preservation
Out-of-Scope Uses
This dataset is not intended to be:
- A complete public leaderboard without an additional private hidden test set
- A substitute for professional legal, medical, or government translation review
- A general-purpose Traditional Chinese corpus
- Evidence that one model is broadly better at all Chinese-language tasks
- A source of real user data
Data Creation and Review Notes
This is a synthetic benchmark intended for evaluation pipeline development, preliminary model comparison, error analysis, and rubric design.
The current review marker is:
needs_native_taiwanese_reviewer_pass_before_public_release
For a public leaderboard, the recommended process is:
- Add a Taiwan-native reviewer pass before freezing the references.
- Keep an additional private hidden test set.
- Do not rely on a single reference-matching metric as the final ranking criterion.
Checksums
Checksums for the v2 release files are recorded in full/tw_localization_benchmark_v2_manifest.json.
Citation
@dataset{openformosa_zh_translation_benchmark_2026,
title = {zh_translation_benchmark},
author = {OpenFormosa},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/OpenFormosa/zh_translation_benchmark}
}
License
Released under the Creative Commons Attribution 4.0 International license (cc-by-4.0) unless superseded by a later repository-level license notice.