LiveCodeBench

Official repository for the paper "LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code"

🏠 Home Page • 💻 Data • 🏆 Leaderboard

Introduction

LiveCodeBench provides holistic and contamination-free evaluation of coding capabilities of LLMs. Particularly, LiveCodeBench continuously collects new problems over time from contests across three competition platforms -- LeetCode, AtCoder, and CodeForces. Next, LiveCodeBench also focuses on a broader range of code-related capabilities, such as self-repair, code execution, and test output prediction, beyond just code generation. Currently, LiveCodeBench hosts four hundred high-quality coding problems that were published between May 2023 and March 2024.

Installation

You can clone the repository using the following command:

git clone https://github.com/LiveCodeBench/LiveCodeBench.git
cd LiveCodeBench

We recommend using poetry for managing dependencies. You can install poetry and the dependencies using the following commands:

pip install poetry
poetry install

The default setup does not install vllm. To install vllm as well you can use:

poetry install --with with-gpu

Data

We provide a benchmark for different code capability scenarios

• Code Generation
• Code Execution
• Test Output Prediction

Inference and Evaluation

Code Generation

We use vllm for inference using open models. By default, we use tensor_parallel_size=${num_gpus} to parallelize inference across all available GPUs. It can be configued using the --tensor_parallel_size flag as required.

For running the inference, please provide the model_name based on the ./lcb_runner/lm_styles.py file. The scenario (here codegeneration) can be used to specify the scenario for the model.

python -m lcb_runner.runner.main --model {model_name} --scenario codegeneration

Additionally, --use_cache flag can be used to cache the generated outputs and --continue_existing flag can be used to use the existing dumped results. In case you wish to use model from a local path, you can additionally provide --local_model_path flag with the path to the model. We use n=10 and temperature=0.2 for generation. Please check the ./lcb_runner/runner/parser.py file for more details on the flags.

For closed API models, --multiprocess flag can be used to parallelize queries to API servers (adjustable according to rate limits).

Evaluation

We compute pass@1 and pass@5 metrics for model evaluations. We use a modified version of the checker released with the apps benchmark to compute the metrics. Particularly, we identified some unhandled edge cases in the original checker and fixed them and additionally simplified the checker based on our collected dataset. To run the evaluation, you can add the --evaluate flag:

python -m lcb_runner.runner.main --model {model_name} --scenario codegeneration --evaluate

Note that time limits can cause slight (< 0.5) points of variation in the computation of the pass@1 and pass@5 metrics. If you observe a significant variation in performance, adjust the --num_process_evaluate flag to a lower value or increase the --timeout flag. Please report particular issues caused by improper timeouts here.

Finally, to get scores over different time windows, you can use ./lcb_runner/evaluation/compute_scores.py file. Particularly, you can provide --start_date and --end_date flags (using the YYYY-MM-DD format) to get scores over the specified time window. In our paper, to counter contamination in the DeepSeek models, we only report results on problems released after August 2023. You can replicate those evaluations using:

python -m lcb_runner.evaluation.compute_scores --eval_all_file {saved_eval_all_file} --start_date 2023-09-01

NOTE: We have pruned a large number of test cases from the original benchmark and created code_generation_lite which is set as the default benchmark offering similar performance estimation much faster. If you wish to use the original benchmark, please use the --not_fast flag. We are in the process of updating the leaderboard scores with this updated setting.

NOTE: V2 Update: to run the update LiveCodeBench please use --release_version release_v2. In addition, if you have existing results from release_v1 you can add --continue_existing or better --continue_existing_with_eval flags to reuse the old completions or evaluations respectively.

Self Repair

For running self repair, you need to provide an additional --codegen_n flag that maps to the number of codes that were generated during code generation. Additionally, the --temperature flag is used to resolve the old code generation eval file which must be present in the output directory.

python -m lcb_runner.runner.main --model {model_name --scenario selfrepair --codegen_n {num_codes_codegen} --n 1 # only n=1 supported

In case you have results on a smaller subset or version of the benchmark, you can use --continue_existing and --continue_existing_with_eval flags to reuse the old computations. Particularly, you can run the following command to continue from existing generated solutions.

python -m lcb_runner.runner.main --model {model_name} --scenario selfrepair --evaluate --continue_existing

Note that this will only reuse the generated samples and rerun evaluations. To reuse the old evaluations, you can add the --continue_existing_with_eval flag.

Test Output Prediction

For running the test output prediction scenario you can simply run

python -m lcb_runner.runner.main --model {model_name} --scenario testoutputprediction --evaluate

Code Execution

For running the test output prediction scenario you can simply run

python -m lcb_runner.runner.main --model {model_name} --scenario codeexecution --evaluate

Additionally, we support the COT setting with

python -m lcb_runner.runner.main --model {model_name} --scenario codeexecution --cot_code_execution --evaluate

Custom Evaluation

Alternatively, you can using lcb_runner/runner/custom_evaluator.py to directly evaluated model generations in a custom file. The file should contain a list of model outputs, appropirately formatted for evaluation in the order of benchmark problems.

python -m lcb_runner.runner.custom_evaluator --custom_output_file {path_to_custom_outputs}

Particularly, arrange the outputs in the following format

[
    {"question_id": "id1", "code_list": ["code1", "code2"]},
    {"question_id": "id2", "code_list": ["code1", "code2"]}
]

Adding Support for New Models

To add support for new models, we have implemented an extensible framework to add new models and customize prompts appropirately.

Step 1: Add a new model to the ./lcb_runner/lm_styles.py file. Particularly, extend the LMStyle class to add a new model family and extend the model to the LanguageModelList array.

Step 2: Since we use instruction tuned models, we allow configuring the instruction for each model. Modify the ./lcb_runner/prompts/generation.py file to add a new prompt for the model in the format_prompt_generation function. For example, the prompt for DeepSeekCodeInstruct family of models looks as follows

# ./lcb_runner/prompts/generation.py
if LanguageModelStyle == LMStyle.DeepSeekCodeInstruct:
    prompt = f"{PromptConstants.SYSTEM_MESSAGE_DEEPSEEK}\n\n"
    prompt += f"{get_deepseekcode_question_template_answer(question)}"
    return prompt

Submit Models to Leaderboard

To submit models to the leaderboard you can fill out this form. You will need to fill out model details and provide the generated evaluation file with model generations and pass@1 scores. We will review the submission and add the model to the leaderboard accordingly.

Results

LiveCodeBench can be used to evaluate performance of LLMs on different time-windows (using problem release date to filter the models). Thus we can detect and prevent potential contamination in the evaluation process and evaluate LLMs on new problems.

Next, we evaluate models on different code capabilities and find that relative performances of models do change over tasks (left). Thus, it highlights the need for holistic evaluation of LLMs for code.

We also find evidence of possible overfitting on HumanEval (right). Particularly, models that perform well on HumanEval do not necessarily perform well on LiveCodeBench. In the scatterplot above, we find the models get clustered into two groups, shaded in red and green. The red group contains models that perform well on HumanEval but poorly on LiveCodeBench, while the green group contains models that perform well on both.

For more details, please refer to our website at livecodebench.github.io.

Citation

@article{jain2024livecodebench,
  author    = {Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, Ion Stoica},
  title     = {LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code},
  year      = {2024},
  journal   = {arXiv preprint},
}