MHPP is a code genration benchmark that:

• ✨ Assesses LLMs' comprehension of specifications and restrictions, multi-step reasoning, and effective application of coding knowledge.
• ✨ Contains 140 manually curated hard Python problems across 7 different challenge types.

Why MHPP?

• ✨ Precise evaluation & Ranking: See the MHPP Leaderboard. The results align well with HumanEval and also suggest new discoveries.
• ✨ Comprehensive Coverage: Our benchmark includes 7 challenge types, enabling a more rigorous and nuanced evaluation of models' coding capabilities across various difficulties, granularities, and reasoning levels.
• ✨ Data Integrity: We will never release answers and full test units to avoid data contamination, 'cheat' by memorizing test data is impossible.

The dataset can be found in the data directory, which includes details of each question, such as the question's challenge type. The raw questions can be accessed using the 'question' field. We also provide a version of the prompt with a prefix for users, as mentioned in the paper. Creating this dataset was labor-intensive. If you find any errors, please submit an issue to provide feedback. Thank you very much! 🙏

1. Run your model on our dataset in the /data/ folder. You can either take the "prompt" field as input (as we do in the paper) or create a new input by adding more user instructions before or after the "question" field. 💻
2. Save the result in JSONL format. Each line should at least contain 'function_name', 'prompt', 'difficulty_types', and 'response'.
3. Run command below to upload your JSONL file to our server, remember to replace "file_name" to your real filename:

curl -F "file=@file_name.jsonl" https://52.155.102.219:3000/upload
# Example
# curl -F "file=@gpt4o_2024_05_13.jsonl" https://52.155.102.219:3000/upload

4. Wait. ⏳
5. Get your model result with a detailed report. 📊
6. (Optional) If you wish to have your result featured on the leaderboard, or if you have any further questions, please click "File a request" on the leaderboard page. This will allow you to create an issue, fill in the necessary information, and submit it. 📤

Thank you for your interest and participation! 😊

If you use MHPP in your research, please consider citing us:

@article{dai2024mhpp,
  title={MHPP: Exploring the Capabilities and Limitations of Language Models Beyond Basic Code Generation},
  author={Dai, Jianbo and Lu, Jianqiao and Feng, Yunlong and Ruan, Rongju and Cheng, Ming and Tan, Haochen and Guo, Zhijiang},
  journal={arXiv preprint arXiv:2405.11430},
  year={2024}
}