README.md

metadata

license: cc-by-4.0
task_categories:
  - translation
language:
  - zh
  - en
size_categories:
  - n<1K

DiscoX Translation Benchmark

DiscoX is a benchmark for the evaluation of LLMs on discourse- and expert-level translation tasks.

Dataset At A Glance

• Languages: English ⇄ Chinese (100 English→Chinese tasks, 100 Chinese→English tasks)
• Total samples: 200 discourse- and exprt-level translation items
• Average passage length: ~1.7k characters (min 0.73k, max 3.04k)
• Meta fields: primary & secondary domain labels, structured rubrics, prompt IDs,etc
• Reference Rubrics: every task ships with multiple rubrics annotated by experts, capturing key points for evaluating translation quality Primary domain coverage:

  Primary Domain	Samples	Share
  学术论文 (Academic papers)	121	60.5%
  非学术论文 (Non-Academic tasks)	79	39.5%

Secondary domain highlights include Social Scienices(社会科学),Natural Sciences(自然科学),Humanities(人文科学),Applied Disciplines(应用学科),News&Information(新闻资讯),Domain-Specific Scenarios(垂类场景) and Literature&Arts(文学艺术).

File Structure

• discox.json: the core dataset. Each record contains
  • ori_text: the source text to be translated
  • prompt: text adding translation instructions
  • reference_list: rubrics designed for evaluating translation results
  • Primary_Domain, Secondary_Domain: high-level topic labels
  • prompt_id, __internal_uuid__: identifiers for specific tasks

Notes & Recommendations

• The reference_list entries are designed to enable targeted verification of translation fidelity: by converting them into structured checks (e.g., terminology, tone, and named entities), the evaluation performs fine-grained, pointwise assessments of key translation aspects.
• Translation instruction in pormpt describe desired output language in Chinese.

License

Our data is under cc-by-4.0 license.