MultiPL-E is a multi-programming language benchmark for evaluating the code generation performance of large language model (LLMs) of code.
We use a suite of compilers to translate the Python benchmarks from Chen et al. 2021 into parallel benchmarks in 18 programming languages. Thus MultiPL-E provides a way to evaluate code generation models on a consistent set of benchmark problems across many languages. The 18 languages capture a broad spectrum of language features, application areas, and popularity, allowing us to explore the impact of these factors on model performance.
Here are some examples of translated problems and Codex-generated solutions in a selection of languages:
The following graph shows the success rate of OpenAI Codex on the complete benchmark across the full suite of programming languages:
Our paper also reports the performance on InCoder, and presents more in depth analyses. Some of our findings include:
Model perplexity is not strongly correlated with the correctness of generated code.
Type annotations have limited impact on model performance for gradually typed languages.
Model performance is correlated with language popularity, but some niche languages perform as well as more popular languages.
The MultiPL-E benchmark and associated tools are open source and easy to extend. Each compiler translates Python unit tests, doctests, and function signatures to its target language. Because these program regions do not contain arbitrary Python code, each compiler is much simpler than a full-fledged compiler. It is easy to add new benchmark programs and to extend the benchmark to additional languages. The tutorial walks you through running our benchmarks and gives guidance on adding support for new languages, new benchmarks, and evaluating new models.