README.md

JumpCoder

📃 Paper | 🔍 About | 🔨 Setup | 🚀 Quick Start | 📊 Evaluation | 🔗 Citation

🔥 News: our paper has been accepted by ACL 2024 as a main conference paper.

🔍 About

This repository contains the implementation of our paper JumpCoder: Go Beyond Autoregressive Coder via Online Modification. JumpCoder is a model-agnostic framework, designed to enhance code large language models (LLMs) through online modification and non-sequential code generation. It does not require any fine-tuning or human intervention.

Why JumpCoder? Traditional code LLMs generate code in a linear, irreversible sequence. This can lead to errors accumulating over time, as they can't jump back and correct the previous missing statements like a human. JumpCoder addresses this issue and improves the quality of the generated code.

How it Works: JumpCoder can insert new code into currently generated code on-the-fly. It uses an auxiliary infilling model that operates alongside the code LLM, allowing for online modifications.

🔨 Setup

Important

JumpCoder requires separate environments for the infilling and generation models, as they might need different versions of libraries like transformers. These models communicate through socket connections. Be sure to set up each model in its appropriate environment.

For example, if using CodeLlama, ensure you have transformers version 4.33.0 or newer. Refer to official instructions for setup.

For WizardCoder, use transformers==4.31.0. Setup instructions can be found here.

🚀 Quick Start

To use JumpCoder, you'll need an infilling model. We recommend CodeLlama-Instruct-7B, which performs well in our experiments.

# in the CodeLlama environment
python run_llm_server.py \
    --checkpoint codellama/CodeLlama-7b-Instruct-hf \
    --port 12345 \
    --device cuda:0

Next, launch the generation model. For example, using CodeLlama-Instruct-7B as well:

# in the CodeLlama environment
python run_llm_server.py \
    --checkpoint codellama/CodeLlama-7b-Instruct-hf \
    --port 12346 \
    --device cuda:1

To test the example, run:

python example.py --port_infilling 12345 --port_generation 12346

You can observe both the standard autoregression output and our JumpCoder's output, which includes dynamic code infilling.

Click here for the example output

# Generated from Autoregressive Coder
def minSubArraySum(nums):
    """
    Given an array of integers nums, find the minimum sum of any non-empty sub-array
    of nums.
    Example
    minSubArraySum([2, 3, 4, 1, 2, 4]) == 1
    minSubArraySum([-1, -2, -3]) == -6
    """
    if not nums:
        return 0
    min_sum = sum(nums)
    for i in range(len(nums)):
        for j in range(i, len(nums)):
            if sum(nums[i:j]) < min_sum:
                min_sum = sum(nums[i:j])
    return min_sum

# Generated from JumpCoder
def minSubArraySum(nums):
    """
    Given an array of integers nums, find the minimum sum of any non-empty sub-array
    of nums.
    Example
    minSubArraySum([2, 3, 4, 1, 2, 4]) == 1
    minSubArraySum([-1, -2, -3]) == -6
    """
    if not nums:
        return 0
    min_sum = sum(nums)
    for i in range(len(nums)):
        curr_sum = 0  # infill
        for j in range(i, len(nums)):
            curr_sum += nums[j]
            if curr_sum < min_sum:
                min_sum = curr_sum
    return min_sum

📊 Evaluation

We tested JumpCoder using several CodeLlama and WizardCoder variants on HumanEval and MBPP.

To demonstrate its versatility in different programming languages, we also evaluated it using the MultiPL-E benchmark.

You can find our test scripts in the scripts/ directory. Below is an example command for evaluating CodeLlama-7B-Instruct on HumanEval. Adjust the parameters as needed (model path, dataset path, etc.).

# Launch the generation server
bash scripts/evaluate_cl_humaneval.sh --build-generation-server

# In a new terminal, launch the infilling server
bash scripts/evaluate_cl_humaneval.sh --build-infilling-server

# Finally, run the evaluation
bash scripts/evaluate_cl_humaneval.sh --evaluate

Tip

When testing WizardCoder, make sure to execute --build-generation-server and --build-infilling-server in their respective Python environments as per their specific requirements.

🔗 Citation

@misc{chen2024jumpcoder,
      title={JumpCoder: Go Beyond Autoregressive Coder via Online Modification}, 
      author={Mouxiang Chen and Hao Tian and Zhongxin Liu and Xiaoxue Ren and Jianling Sun},
      year={2024},
      eprint={2401.07870},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}