EpiCoder-func-380k is a dataset released by Microsoft in 2025, focused on fine-tuning the code generation capabilities of large language models. The related paper is "EpiCoder: Encompassing Diversity and Complexity in Code Generation", and the dataset aims to address challenges related to the diversity and complexity of instructions in code generation tasks.

The dataset contains 380,000 function-level instruction-output pairs, specifically designed to train large language models to generate Python code based on detailed textual instructions. The dataset was constructed using synthetic methods, with both instructions and code generated by large language models and verified for correctness in a Docker environment. It covers a variety of programming scenarios, including instruction-level parallel optimization, and is suitable for enhancing model performance on specific code tasks.

Dataset Composition

The dataset primarily consists of the following two fields, with each entry representing a complete programming task instance:

• instruction: A string type containing a detailed description of the programming task, specifying the functionality, constraints, and optimization requirements for the code.
• output: A string type containing the Python code implementation that satisfies the aforementioned instruction.

All data consists of expert-generated instructions paired with model-synthesized code, and does not include additional metadata or comment fields.

Citation

@misc{wang2025epicoderencompassingdiversitycomplexity,
      title={EpiCoder: Encompassing Diversity and Complexity in Code Generation}, 
      author={Yaoxiang Wang and Haoling Li and Xin Zhang and Jie Wu and Xiao Liu and Wenxiang Hu and Zhongxin Guo and Yangyu Huang and Ying Xin and Yujiu Yang and Jinsong Su and Qi Chen and Scarlett Li},
      year={2025},
      eprint={2501.04694},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2501.04694}, 
}