Home

Awesome

BioCoder

<p align="center"><a href="https://arxiv.org/abs/2308.16458">šŸ“„ Paper</a> <a href="https://biocoder-benchmark.github.io">šŸŒ Website</a> <a href="https://huggingface.co/datasets/lilbillbiscuit/biocoder_public">šŸ“œ Dataset</a>

BioCoder is a challenging bioinformatics code generation benchmark for examining the capabilities of state-of-the-art large language models (LLMs).

Project Structure

The repository comprises 4 main parts: evaluation, inference, parsing, and data. We use the data folder to store results between each step in the operations. Ensure that each step is run from the root directory, i.e. python3 parsing/parse.py and not cd parsing && python3 parse.py in order to avoid issues with relative paths.

Each section is described below.

Parsing

The parsing section contains the code used to analyze, parse, and extract functions from the repositories.

Begin the process by going through a list of GitHub repositories. Download the list of repositories (named pone.0205898.s005.xlsx) from this paper first, then place it in the data directory.

A large-scale analysis of bioinformatics code on GitHub

Then, open the notebook in GithubAnalysis to start the repository metadata-gathering process. Note that this may fully clone all available GitHub repositories in the list. This may take a long time. Then, use the scripts in FunctionSelection to filter the GitHub functions accordingly.

Next, we need to do the function extraction step. We have created parsers for Java and Python so far. The corresponding documentation can be found in the respective folders:

After parsing, the output should be a JSON format with consistent properties across all languages. The following columns are guaranteed, however, they may be extra columns for specific languages:

Then, use the scripts in FunctionProcessing to process the functions. This includes filtering out functions that are too short, too long, or have too many parameters. It also includes filtering out functions that are too similar to each other. It also includes scripts to generate summaries, prompts, and other necessary prompt data with ChatGPT

Note that there are also some scripts that assist with the manual annotation of context generation. These are not necessary for the benchmark, but are included for completeness.

Custom Python Parsing

In order to perform custom parsing on a Python repository, one can use the parse.py file under parsing/FunctionExtractors/parse.py. Before running one must make sure that a json_files folder must be in the same repository as parse.py as well as agithub_repos folder must be in the same repository as parse.py. In the github_repos folder, one can include the physical repositories that one desires to parse. In order to run parse.py one must define the PACKAGE_DIRECTORY, PACKAGE_NAME, and REPO_AUTHOR variables. For instance, if you want to parse the files under the Bio/AlignIO folder in the biopython repository, with the author of the repository also being biopython, then you would set

PACKAGE_DIRECTORY = biopython/Bio/AlignIO
PACKAGE_NAME = AlignIO
REPO_AUTHOR = biopython

and here is some sample code to run the parser:

import os, subprocess

directory = 'biopython/Bio/AlignIO'
author = 'biopython'

os.environ['PACKAGE_DIRECTORY'] = os.path.join('./github_repos', directory)
package_name = directory.split('/')[-1]
os.environ['PACKAGE_NAME'] = package_name
os.environ['REPO_AUTHOR'] = author
subprocess.run('python3 parse.py', shell=True)

The result will be a file named {PACKAGE_NAME}_functions.json which includes the main json file containing all the parsed functions. There will also be a file named AlignIO_file_imports.json which includes the import packages necessary to complete the parsing.

Loading the dataset

The dataset is located in the datasets folder. Most of the following steps require the data to be structured as follows:

/data/Prompts/
- [Prompt Type 1]
  - [Prompt ID 1]
  - [Prompt ID 2]
  - ...
  - [Prompt ID N]
- [Prompt Type 2]
    - [Prompt ID 1]
    - [Prompt ID 2]
    - ...
    - [Prompt ID N]
- ...

We have already parsed the public dataset in this format. Just extract it with

unzip datasets/Prompts.zip -d data/

Inference

This section consists of all files necessary to generate the outputs across different models.

For convenience on running the script across multiple machines (output generation takes a long time, especially for these prompts), we have included a requirements.txt with the exact dependencies that we used.

The following arguments are available:

After installing the dependencies, you can install everything with pip3 install -r requirements.txt.

Then, run the following command to generate the outputs:

python3 inference/final_batch_run.py --num_gpus 8 --gpus_per_process 1 --model_name_or_path bigcode/starcoder --generation_version 1 --max_length 8192 --max_generation 200 --PROMPT_AMOUNT 20 --use_summary_only True --tolerance 0.9 --base_folder_url data/ --prompt_basefolder prompts

Evaluation

Setting up the evaluation framework

The evaluation framework utilizes a combination of Lambda and SQS to assign tasks across multiple workers. The workers are Docker containers that are spun up on demand. The Docker containers are spun up repeatedly and query the Lambda function for tasks to perform. DynamoDB is used to store the results.

We will publish a version that runs completely locally in the future.

  1. Create an AWS account
  2. Create an AWS Lambda function with the following parameters:

  1. Copy the contents of CentralServer/lambda_queue_handler.py into the Lambda function

  2. Click Deploy

  3. Create an EC2 Launch Template with the following settings:

  4. Create two SQS queues, for example, one named bioqueue and one named bioqueue-inprogress

  5. Create an S3 bucket, for example, named biodata (names are globally unique so these names might not be available)

  6. Create a DynamoDB table with the following settings:

Load test cases with the following command:

PROD=true python evaluation/TestCaseLoader/upload_test_cases.py

Create a Docker image with evaluation/DockerServers/Dockerfile

Setting up Rosalind Evaluation Framework

  1. First do all the Rosalind evaluations in the folder named "TestRosalind", and ensure that the following files are available: "run_script.sh", "skeleton.py", and "tester.py"
  2. Make sure the code generation outputs are in a folder called "outputs" with the following structure: under "outputs" there is a folder for each Rosalind problem, i.e. there will be a folder named "2sat", a folder named "2sum", etc. Under each of these folders there will be 20 code generation text files, numbered "1.txt" through "20.txt". Here is a visual representation

outputs/<br> ā”œā”€ 2sat/<br> ā”‚ Ā ā”œā”€ 1.txt<br> ā”‚ Ā ā”œā”€ 2.txt<br> ā”‚ Ā ...<br> ā”‚ Ā ā”œā”€ 20.txt<br> ā”œā”€ 2sum/<br> ā”œā”€ afrq/<br> ,,,

  1. When the "outputs" folder is currently formated as explained above, you can run the "tester.py" file under "TestRosalind". After this file finishes running, there will be a "results.json" file that stores a spreadsheet of the results of the code generation outputs. There are three columns: "Problem", "Generation", and "Passed". "Problem" indicates the name of the problem that was tested, "Generation" indicates the generation number, and "Passed" will be a boolean of whether the specified generation number of the specified problem actually resulted in accurate code. For instance if a row in the spreadsheet is ["2sat", 3, True], then this means that the generated code in "TestRosalind/outputs/2sat/3.txt" is correct and passes the problem.

  2. Here is some sample code that can be used to extract Pass@K results from the "results.json" file

import numpy as np
import pandas as pd
import os

def pass_at_k(n, c, k):
    if n - c < k: return 1.0
    return 1.0 - np.prod(1.0 - k / np.arange(n - c + 1, n + 1))

df = pd.read_json('results.json')

pk_df = pd.DataFrame(columns=['Problem', 'NumPassed', 'PassAt1', 'PassAt5', 'PassAt10', 'PassAt20'])

num_passed = {}
for prob in os.listdir('./TestRosalind/outputs'):
    num_passed[prob] = 0

for _, row in df.iterrows():
    if row['Passed']:
        num_passed[row['Problem']] += 1

for prob in num_passed.keys():
    c = num_passed[prob]
    pk_df.loc[len(pk_df)] = [prob, c, pass_at_k(20, c, 1), pass_at_k(20, c, 5), pass_at_k(20, c, 10), pass_at_k(20, c, 20)]

print(pk_df['PassAt1'].mean())
print(pk_df['PassAt5'].mean())
print(pk_df['PassAt10'].mean())
print(pk_df['PassAt20'].mean())

Datasets

We provide our completed public dataset in this GitHub repository. It consists of the following files:

Citation

If you find our work useful in your research, please kindly consider cite:

@article{10.1093/bioinformatics/btae230,
    author = {Tang, Xiangru and Qian, Bill and Gao, Rick and Chen, Jiakang and Chen, Xinyun and Gerstein, Mark B},
    title = "{BioCoder: a benchmark for bioinformatics code generation with large language models}",
    journal = {Bioinformatics},
    volume = {40},
    number = {Supplement_1},
    pages = {i266-i276},
    year = {2024},
    month = {06},
    abstract = "{Pretrained large language models (LLMs) have significantly improved code generation. As these models scale up, there is an increasing need for the output to handle more intricate tasks and to be appropriately specialized to particular domains. Here, we target bioinformatics due to the amount of domain knowledge, algorithms, and data operations this discipline requires. We present BioCoder, a benchmark developed to evaluate LLMs in generating bioinformatics-specific code. BioCoder spans much of the field, covering cross-file dependencies, class declarations, and global variables. It incorporates 1026 Python functions and 1243 Java methods extracted from GitHub, along with 253 examples from the Rosalind Project, all pertaining to bioinformatics. Using topic modeling, we show that the overall coverage of the included code is representative of the full spectrum of bioinformatics calculations. BioCoder incorporates a fuzz-testing framework for evaluation. We have applied it to evaluate various models including InCoder, CodeGen, CodeGen2, SantaCoder, StarCoder, StarCoder+, InstructCodeT5+, GPT-3.5, and GPT-4. Furthermore, we fine-tuned one model (StarCoder), demonstrating that our training dataset can enhance the performance on our testing benchmark (by \\&gt;15\\% in terms of Pass@K under certain prompt configurations and always \\&gt;3\\%). The results highlight two key aspects of successful models: (i) Successful models accommodate a long prompt (\\&gt;2600 tokens) with full context, including functional dependencies. (ii) They contain domain-specific knowledge of bioinformatics, beyond just general coding capability. This is evident from the performance gain of GPT-3.5/4 compared to the smaller models on our benchmark (50\\% versus up to 25\\%).All datasets, benchmark, Docker images, and scripts required for testing are available at: https://github.com/gersteinlab/biocoder and https://biocoder-benchmark.github.io/.}",
    issn = {1367-4811},
    doi = {10.1093/bioinformatics/btae230},
    url = {https://doi.org/10.1093/bioinformatics/btae230},
    eprint = {https://academic.oup.com/bioinformatics/article-pdf/40/Supplement\_1/i266/58354818/btae230.pdf},
}