This repository contains code to perform execution-based multi-lingual evaluation of code generation capabilities and the corresponding data, namely, a multi-lingual benchmark MBXP, multi-lingual MathQA and multi-lingual HumanEval. Results and findings can be found in the paper "Multi-lingual Evaluation of Code Generation Models" (https://arxiv.org/abs/2210.14868).
Our paper describes the language conversion framework, the synthetic solution generation, and many other types of evaluation beyond the traditional function completion evaluation such as translation, code insertion, summarization, and robustness evaluation.
Below we demonstrate the language conversion (component A above) for the conversion from Python to Java (abridged example for brevity).
Check out and install this repository:
git clone https://github.com/amazon-science/mxeval.git
pip install -e mxeval
We provide scripts to help set up programming language dependencies that are used to execute and evaluate using datasets in MBXP.
bash language_setup/amazon_linux_ami.sh
bash language_setup/ubuntu.sh
This program exists to run untrusted model-generated code. Users are strongly
encouraged not to do so outside of a robust security sandbox. See the comment in
execution.py for more information and instructions.
Each sample is formatted into a single line:
{"task_id": "Corresponding task ID", "completion": "Completion only without the prompt",
"language": "programming language name"}
We provide data/mbxp/examples/mbxp_samples.jsonl to illustrate the format.
Here is nearly functional example code (you just have to provide
generate_one_completion to make it work) that saves generated completions to
samples.jsonl.
from mxeval.data import write_jsonl, read_problems
problems = read_problems()
num_samples_per_task = 200
samples = [
dict(task_id=task_id, language=problems[task_id]["language"], completion=generate_one_completion(problems[task_id]["prompt"]))
for task_id in problems
for _ in range(num_samples_per_task)
]
write_jsonl("samples.jsonl", samples)
To evaluate the samples for, e.g., Java MBJP evaluation, run
evaluate_functional_correctness data/mbxp/examples/mbjp_samples.jsonl --problem_file data/mbxp/mbjp_release_v1.jsonl
or to run all languages
for lang in mbcpp mbcsp mbgp mbjp mbjsp mbkp mbphp mbplp mbpp mbrbp mbscp mbswp mbtsp;
do evaluate_functional_correctness --problem_file data/mbxp/${lang}_release_v1.jsonl data/mbxp/examples/${lang}_samples.jsonl;
done
You can check the programming-language dependency installation by running the above example for each MBXP dataset. You should obtain the following results for the mbxp_samples.jsonl files provided:
| Dataset | pass@1 |
|---|---|
| MBCPP | 79.60% |
| MBCSP | 63.63% |
| MBGP | 39.19% |
| MBJP | 85.30% |
| MBJSP | 78.67% |
| MBKP | 63.77% |
| MBPHP | 72.77% |
| MBPLP | 38.41% |
| MBPP | 82.24% |
| MBRBP | 58.90% |
| MBSCP | 42.96% |
| MBSWP | 29.40% |
| MBTSP | 87.29% |
Note: Because there is no unbiased way of estimating pass@k when there are fewer
samples than k, the script does not evaluate pass@k for these cases. To
evaluate with other k values, pass --k <comma-separated-values-here>. For
other options, see
$ evaluate_functional_correctness --help
However, we recommend that you use the default values for the rest.
evaluate_functional_correctness data/mbxp/samples/mbjp_samples.jsonl --problem_file data/mbxp/mbjp_release_v1.jsonl --n_workers 63 --k 1,5,10,100
While evaluation uses very little memory, you might see the following error message when the system is running out of RAM. Since this may cause some correct programs to fail, we recommend that you free some memory and try again.
malloc: can't allocate region
Some system might require longer compilation timeout. If you see that the execution fails due to compilation timeout reason, this number should be increased accordingly.
We plan to release synthetic canonical solutions as well as processed datasets for other evaluation tasks such as code-insertion, code-translation, etc.
We adapted OpenAI's human-eval package (https://github.com/openai/human-eval) for the multi-lingual case. We thank OpenAI for their pioneering effort in this field including the release of the original HumanEval dataset, which we convert to the multi-lingual versions. We also thank Google for their release of the original MBPP Python dataset (https://github.com/google-research/google-research/tree/master/mbpp), which we adapt and convert to other programming languages.
Please cite using the following bibtex entry:
@article{mbxp_athiwaratkun2022,
title = {Multi-lingual Evaluation of Code Generation Models},
author = {Athiwaratkun, Ben and
Gouda, Sanjay Krishna and
Wang, Zijian and
Li, Xiaopeng and
Tian, Yuchen and
Tan, Ming
and Ahmad, Wasi Uddin and
Wang, Shiqi and
Sun, Qing and
Shang, Mingyue and
Gonugondla, Sujan Kumar and
Ding, Hantian and
Kumar, Varun and
Fulton, Nathan and
Farahani, Arash and
Jain, Siddhartha and
Giaquinto, Robert and
Qian, Haifeng and
Ramanathan, Murali Krishna and
Nallapati, Ramesh and
Ray, Baishakhi and
Bhatia, Parminder and
Sengupta, Sudipta and
Roth, Dan and
Xiang, Bing},
doi = {10.48550/ARXIV.2210.14868},
url = {https://arxiv.org/abs/2210.14868},
keywords = {Machine Learning (cs.LG), Computation and Language (cs.CL), FOS: Computer and information sciences, FOS: Computer and information sciences},
publisher = {arXiv},
year = {2022},
copyright = {Creative Commons Attribution 4.0 International}
}