This is a framework for the evaluation of code generation models. This work is inspired from EleutherAI/lm-evaluation-harness for evaluating language models in general. We welcome contributions to fix issues, enhance features and add new benchmarks. You can find contribution guides in docs/guide.md and CONTRIBUTING.md and more documentation in docs/README.md.
Below are the features and tasks of this framework:
-
Features:
- Any autoregressive model available on Hugging Face hub can be used, but we recommend using code generation models trained specifically on Code such as SantaCoder, InCoder and CodeGen.
- We provide Multi-GPU text generation with
accelerateand Dockerfiles for evaluating on Docker containers for security and reproducibility.
-
Tasks:
- 7 code generation Python tasks (with unit tests): HumanEval, HumanEval+, InstructHumanEval, APPS, MBPP, MBPP+, and DS-1000 for both completion (left-to-right) and insertion (FIM) mode.
- HumanEvalPack extends HumanEval to 3 scenarios across 6 languages via human translations and was released with OctoPack.
- MultiPL-E evaluation suite (HumanEval translated into 18 programming languages).
- Recode applied to the HumanEval benchmark. It evaluates the robustness of code-generation models.
- Pal Program-aided Language Models evaluation for grade school math problems : GSM8K and GSM-HARD. These problems are solved by generating reasoning chains of text and code.
- Code to text task from CodeXGLUE (zero-shot & fine-tuning) for 6 languages: Python, Go, Ruby, Java, JavaScript and PHP. Documentation translation task from CodeXGLUE.
- CoNaLa for Python code generation (2-shot setting and evaluation with BLEU score).
- Concode for Java code generation (2-shot setting and evaluation with BLEU score).
- 3 multilingual downstream classification tasks: Java Complexity prediction, Java code equivalence prediction, C code defect prediction.
- SantaCoder-FIM for evaluating FIM on Python code using Exact Match. Further details are described in SantaCoder. Includes two tasks:
StarCoderFIM: which uses the default FIM tokens"<fim_prefix>", "<fim_middle>", "<fim_suffix>", andSantaCoderFIM: which uses SantaCoder FIM tokens"<fim-prefix>", "<fim-middle>", "<fim-suffix>"
- Mercury for evaluating computational efficiency of Python code generation.
More details about each task can be found in the documentation in docs/README.md.
git clone https://github.com/bigcode-project/bigcode-evaluation-harness.git
cd bigcode-evaluation-harnessInstall torch based on your device type, and install the other packages using:
pip install -e .
To run the DS-1000 benchmark, additional constraints must be resolved.
# python version must be 3.7.10
pip install -e ".[ds1000]" # installs all additional dependencies except PyTorch
# torch==1.12.1 required. Download version with relevant GPU support etc., e.g.,
pip install torch==1.12.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116
# to suppress any tensorflow optimization warnings,
# precede call to "accelerate launch" with "TF_CPP_MIN_LOG_LEVEL=3"
# on some systems, tensorflow will attempt to allocate all GPU memory
# to its process at import which will raise a CUDA out-of-memory error
# setting "export TF_FORCE_GPU_ALLOW_GROWTH=true" resolves this
Also make sure you have git-lfs installed and are logged in the Hub
huggingface-cli login
We use accelerate to generate code/text in parallel when multiple GPUs are present (multi-GPU mode). You can configure it using:
accelerate configThis evaluation harness can also be used in an evaluation only mode, you can use a Multi-CPU setting. For large models, we recommend specifying the precision of the model using the --precision flag instead of accelerate config to have only one copy of the model in memory. You can also load models in 8bit with the flag --load_in_8bit or 4bit with --load_in_4bit if you have bitsandbytes installed with the required transformers and accelerate versions.
The evaluation part (solutions execution) for MultiPL-E requires extra dependencies for some programming languages, we provide a Dockerfile with all dependencies, see section Docker for more details.
You can use this evaluation harness to generate text solutions to code benchmarks with your model, to evaluate (and execute) the solutions or to do both. While it is better to use GPUs for the generation, the evaluation only requires CPUs. So it might be beneficial to separate these two steps. By default both generation and evaluation are performed.
For more details on how to evaluate on the tasks, please refer to the documentation in docs/README.md.
Below is an example to generate and evaluate on a task.
accelerate launch main.py \
--model <MODEL_NAME> \
--tasks <TASK_NAME> \
--limit <NUMBER_PROBLEMS> \
--max_length_generation <MAX_LENGTH> \
--temperature <TEMPERATURE> \
--do_sample True \
--n_samples 100 \
--batch_size 10 \
--precision <PRECISION> \
--allow_code_execution \
--save_generationslimitrepresents the number of problems to solve, if it's not provided all problems in the benchmark are selected.allow_code_executionis for executing the generated code: it is off by default, read the displayed warning before calling it to enable execution.- Some models with custom code on the HF hub like SantaCoder require calling
--trust_remote_code, for private models add--use_auth_token. save_generationssaves the post-processed generations in a json file atsave_generations_path(by defaultgenerations.json). You can also save references by calling--save_referencesmax_length_generationis the maximum token length of generation including the input token length. The default is 512, but for some tasks like GSM8K and GSM-Hard, the complete prompt with 8 shot examples (as used in PAL) take up~1500tokens, hence the value should be greater than that and the recommended value ofmax_length_generationis2048for these tasks.
Some tasks don't require code execution such as
codexglue_code_to_text-<LANGUAGE>/codexglue_code_to_text-python-left/conala/concode that use BLEU evaluation. In addition, we generate one candidate solution for each problem in these tasks, so use n_samples=1 and batch_size=1. (Note that batch_size should always be equal or less than n_samples).
- For APPS tasks, you can use
n_samples=1for strict and average accuracies (from the original APPS paper) andn_samples>1for pass@k.
If you want to generate solutions without executing and evaluating the code, call --generation_only, in addition to the instructions above. This will save the solutions in a json file provided in save_generation_path in the working directory.
This can be useful if you don't want to execute code in the machine you're using for generations for security or efficiency reasons. For instance, you can do the generations on multiple GPUs, but switch to a multiple workers CPU machine or docker container for the execution.
If you already have the generations in a json file from this evaluation harness and want to evaluate them, specify the path of the generations via the load_generations_path argument. You may need to reconfigure accelerate to use multiple CPUs.
Below is an example, be mind of specifying arguments proper to the task you are evaluating on, and note that model value here only serves for documenting the experiment. Also add --n_samples to specify the number of samples to evaluate per problem (usually the same value used in generation).
accelerate launch main.py --tasks mbpp --allow_code_execution --load_generations_path generations.json --model incoder-temperature-08For safety, we provide a Dockerfiles to do the execution inside a docker container. To do that, first, do the generation on your machine and save them in generations.json for example by adding the flag --generation_only to the command. Then use the Docker image that we provide:
$ docker pull ghcr.io/bigcode-project/evaluation-harness
$ docker tag ghcr.io/bigcode-project/evaluation-harness evaluation-harnessIf you want to evaluate on MultiPL-E, we have a different Dockerfile since it requires more dependencies, use:
$ docker pull ghcr.io/bigcode-project/evaluation-harness-multiple
$ docker tag ghcr.io/bigcode-project/evaluation-harness-multiple evaluation-harness-multipleIf you modify the evaluation harness, you may want to rebuild the docker images.
Here's how to build a docker image for the evaluation harness:
$ sudo make DOCKERFILE=Dockerfile allThis creates an image called evaluation-harness, and runs a test on it. To skip the test remove all form the command.
For MultiPL-E:
$ sudo make DOCKERFILE=Dockerfile-multiple allThis creates an image called evaluation-harness-multiple.
Suppose you generated text with the bigcode/santacoder model and saved it in generations_py.json with:
accelerate launch main.py \
--model bigcode/santacoder \
--tasks multiple-py \
--max_length_generation 650 \
--temperature 0.8 \
--do_sample True \
--n_samples 200 \
--batch_size 200 \
--trust_remote_code \
--generation_only \
--save_generations \
--save_generations_path generations_py.jsonTo run the container (here from image evaluation-harness-multiple) to evaluate on generations_py.json, or another file mount it with -v, specify n_samples and allow code execution with --allow_code_execution (and add the number of problems --limit if it was used during generation):
$ sudo docker run -v $(pwd)/generations_py.json:/app/generations_py.json:ro -it evaluation-harness-multiple python3 main.py \
--model bigcode/santacoder \
--tasks multiple-py \
--load_generations_path /app/generations_py.json \
--allow_code_execution \
--temperature 0.8 \
--n_samples 200To implement a new task in this evaluation harness, see the guide in docs/guide. The are also contribution guidelines in this CONTRIBUTING.md
We provide documentation for the existing benchmarks and how to run the evaluation in docs/README.md.
- Currenltly, we use data parallel evaluation across multiple GPUs using
accelerate, this assumes that you can fit the model in one GPU.
We thank EleutherAI for their work on the lm-evaluation harness from which this repository is inspired.
@misc{bigcode-evaluation-harness,
author = {Ben Allal, Loubna and
Muennighoff, Niklas and
Kumar Umapathi, Logesh and
Lipkin, Ben and
von Werra, Leandro},
title = {A framework for the evaluation of code generation models},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/bigcode-project/bigcode-evaluation-harness}},
year = 2022,
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent