yeqing0218 / codefuse-evaluation

CodeFuseEval: Multi-tasking Evaluation Benchmark for Code Large Language Model

中文 ｜ English

CodeFuseEval is a Code Generation benchmark that combines the multi-tasking scenarios of CodeFuse Model with the benchmarks of HumanEval-x and MBPP. This benchmark is designed to evaluate the performance of models in various multi-tasking tasks, including code completion, code generation from natural language, test case generation, cross-language code translation, and code generation from Chinese commands, among others.

Generation environment：

CodeFuse-13B: Python 3.8 or above,PyTorch 1.12 or above, with a recommendation for 2.0 or above, Transformers 4.24.0 or above ,CUDA 11.4 or above (for GPU users and flash-attention users, this option should be considered).

CodeFuse-CodeLlama-34B:python>=3.8,pytorch>=2.0.0,transformers==4.32.0,Sentencepiece,CUDA 11.

Generation Comand：

bash codefuseEval/script/generation.sh MODELNAME EVALDATASET OUTFILE LANGUAGE

eg:
bash codefuseEval/script/generation.sh CodeFuse-13B humaneval_python result/test.jsonl python

How to use codefuseEval

Evaluation Data

Data are stored in codefuseEval/data, using JSON list format. We first integrated humaneval-X dataset.

• task_id: indicates the target language and ID of the problem. Language is one of ["Python", "Java", "JavaScript", "CPP", "Go"].
• prompt: the function declaration and docstring, used for code generation.
• declaration: only the function declaration, used for code translation.
• canonical_solution: human-crafted example solutions.
• test: hidden test samples, used for evaluation
• example_test: public test samples (appeared in prompt), used for evaluation.
• prompt_text: prompt text
• prompt_explain: prompt explanation
• func_title: code function title
• prompt_text_chinese: Chinese prompt

Evaluation Environment

The evaluation of the generated codes involves compiling and running in multiple programming languages. The versions of the programming language environments and packages we use are as follows:

Dependency	Version
Python	3.8.12
JDK	18.0.2.1
Node.js	16.14.0
js-md5	0.7.3
C++	11
g++	7.5.0
Boost	1.71.0
OpenSSL	3.0.0
go	1.18.4

In order to save everyone the trouble of setting up the environments for these languages, we create a Docker image with the required environments and codefuseEval.

docker pull registry.cn-hangzhou.aliyuncs.com/codefuse/codefuseeval:latest

If you are familiar with docker, you can build the image from codefuseEval/docker/Dockerfile or configure the Dockerfile as you like it:

cd codefuseEval/docker
docker build [OPTIONS] .

After obtaining the image, you can build a container using the following command:

docker run -it --gpus all --mount type=bind,source=<LOCAL PATH>,target=<PATH IN CONTAINER> [OPTIONS] <IMAGE NAME:TAG>

Evaluation Metrics

In addition to the unbiased pass@k indicators currently provided in Codex, we will also integrate the relevant indicators of huggingface open source with CodeBLEU for integration. The main indicators currently recommended for users are as follows:

• codebleu
• pass@k
• bleu
• bleurt

For other related metrics, you can check the code of the metric or the evaluation code to meet your requirements.

Evaluation

We recommend evaluating in the provided image. To evaluate the generated samples, save generated codes in the following JSON list format:

{"task_id": "../..", "generation: "..."}
{"task_id": "../..", "generation: "..."}
...

and evaluate them using the following script under the root directory of the repository (please execute with caution, the generated codes might have unexpected behaviours though with very low possibility. See the warnings in execution.py and uncomment the execution lines at your own risk):

bash codefuseEval/script/evaluation.sh <RESULT_FILE> <METRIC> <PROBLEM_FILE> <TEST_GROUDTRUTH>
eg: 
bash codefuseEval/script/evaluation.sh codefuseEval/result/test.jsonl pass@k humaneval_python

At the same time, we currently provide the following flags, which can directly bring the sample answers in the test data set as generated answers for testing.

• TEST_GROUDTRUTH default False

When TEST_GROUDTRUTH is True, the self-test mode is turned on, PROBLEM_FILE will be read, and the sample answer will be substituted as the generated answer for testing.

When TEST_GROUDTRUTH is False, open the evaluation mode, read RESULT_FILE and PROBLEM_FILE, and substitute the generated answer for testing.

Check result Command：

We provide the script to check the result for provided code LLMs. Please use following scripts to check corresponding results and the environment .

bash codefuseEval/script/check_reference.sh codefuseEval/result/CodeFuse-CodeLlama-34B/humaneval_result_python.jsonl humaneval_python
bash codefuseEval/script/check_reference.sh codefuseEval/result/CodeFuse-13B/humaneval_result_python.jsonl humaneval_python 

Check dataset Command：

bash codefuseEval/script/check_dataset.sh humaneval_python

bash codefuseEval/script/check_dataset.sh humaneval_java

bash codefuseEval/script/check_dataset.sh humaneval_js

bash codefuseEval/script/check_dataset.sh humaneval_rust

bash codefuseEval/script/check_dataset.sh humaneval_go

bash codefuseEval/script/check_dataset.sh humaneval_cpp