README.md

metadata

license: cc-by-4.0
language:
  - en
  - tr
task_categories:
  - text-generation
tags:
  - code
  - code-generation
  - multilingual
  - turkish
  - benchmark
pretty_name: EN-TR CodeBench
viewer: false

EN-TR CodeBench

EN-TR CodeBench is a parallel English–Turkish benchmark for measuring how prompt language affects the code generation performance of large language models.

The dataset will be released upon the publication of the accompanying paper.

Overview

• 60 Python programming problems (15 easy, 20 medium, 25 hard)
• 811 unit tests
• Three evaluation conditions:
  • English: English prompt with English function and parameter names
  • Turkish: Turkish prompt with English function and parameter names
  • Turkish-localized: Turkish prompt with Turkish function and parameter names

Contact

Mert İncidelen — mincidelen@firat.edu.tr