---
license: mit
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
dataset_info:
  features:
  - name: task_id
    dtype: string
  - name: scenario_id
    dtype: string
  - name: env_id
    dtype: string
  - name: api_specification
    dtype: string
  - name: text_specification
    dtype: string
  - name: short_app_description
    dtype: string
  - name: scenario_instructions
    dtype: string
  - name: needs_db
    dtype: bool
  - name: needs_secret
    dtype: bool
  - name: needed_packages
    struct:
    - name: JavaScript
      sequence: string
    - name: _all_
      sequence: string
  - name: potential_cwes
    sequence: int64
  - name: env_language
    dtype: string
  - name: env_extension
    dtype: string
  - name: env_framework
    dtype: string
  - name: env_multifile
    dtype: bool
  - name: code_filename
    dtype: string
  - name: entrypoint_cmd
    dtype: string
  - name: allowed_packages
    dtype: string
  - name: env_instructions
    dtype: string
  - name: port
    dtype: int64
  splits:
  - name: test
    num_bytes: 1830262
    num_examples: 392
  download_size: 70540
  dataset_size: 1830262
task_categories:
- text-generation
tags:
- code
- security
- benchmark
size_categories:
- n<1K
---

### Dataset Summary
BaxBench is a coding benchmark constructed to measure the ability of code generation models and agents to generate correct and secure code. It consists of 392 backend development tasks, which are constructed by combining 28 scenarios that describe the backend functionalities to implement and 14 backend frameworks defining the implementation tools. To assess the correctness and security of the solutions, the benchmark uses end-to-end functional tests and practical securtiy exploits.

The dataset is released as part of the paper and benchmark: [BaxBench: Can LLMs generate Correct and Secure Backends?](https://arxiv.org/abs/2502.11844).

The dataset contains all necessary artifacts to reproduce the evaluation prompts used in our paper. Further, it enables the testing of different prompt structures or models by forming new prompt types, e.g., for testing code agents.

For details on reproducing our results, or testing your models on the same prompts, please refer to our [paper](https://arxiv.org/abs/2502.11844) or [code repository](https://github.com/logic-star-ai/baxbench).

To test your generated solutions, please follow the instructions in our [code repository](https://github.com/logic-star-ai/baxbench).

For more details on the construction of BaxBench, large-scale model evaluation results, and detailed analyses, please see our [paper](https://arxiv.org/abs/2502.11844) or visit our [website](https://baxbench.com).

### Citation

**BibTeX:**
```
@article{vero2025baxbenchllmsgeneratecorrect,
    title={BaxBench: Can LLMs Generate Correct and Secure Backends?},
    author={Mark Vero and Niels Mündler and Victor Chibotaru and Veselin Raychev and Maximilian Baader and Nikola Jovanović and Jingxuan He and Martin Vechev},
    year={2025},
    eprint={2502.11844},
    archivePrefix={arXiv},
}
```