CodeMMLU
收藏资源简介:
CodeMMLU是由FPT软件AI中心和河内科技大学联合创建的一个多任务基准数据集,旨在评估代码大语言模型(CodeLLMs)的代码理解能力。该数据集包含超过10,000个问题,涵盖代码分析、缺陷检测和软件工程原则等多个领域,涉及多种编程语言。数据集的创建过程结合了高质量的编程相关MCQs,并通过深度学习模型进行过滤,确保问题的高质量和多样性。CodeMMLU的应用领域主要集中在提升AI辅助软件开发的可靠性和能力,旨在解决代码生成模型在理解和推理代码方面的不足。
CodeMMLU is a multi-task benchmark dataset jointly created by FPT Software AI Center and Hanoi University of Science and Technology, aiming to evaluate the code comprehension capabilities of code large language models (CodeLLMs). This dataset contains over 10,000 questions covering multiple domains such as code analysis, defect detection and software engineering principles, involving a variety of programming languages. The construction of CodeMMLU combines high-quality programming-related multiple-choice questions (MCQs) and filters them via deep learning models to ensure the high quality and diversity of the questions. The main application scenarios of CodeMMLU focus on enhancing the reliability and capabilities of AI-assisted software development, aiming to address the shortcomings of code generation models in code understanding and reasoning.
CodeMMLU
概述
CodeMMLU 是一个综合基准,旨在评估大型语言模型(LLMs)在编码和软件知识方面的能力。它基于多项选择题(MCQA)的结构,涵盖了广泛的编程任务和领域,包括代码生成、缺陷检测、软件工程原理等。
数据集结构
| 主题 | 子主题 | 大小 |
|---|---|---|
| 语法知识 | API & 框架使用 | 740 |
| 编程语言语法 | 6220 | |
| 语义知识 | DBMS & SQL | 393 |
| 软件原则 | 3246 | |
| 其他 | 程序可访问性、计算机网络等 | 1308 |
| 实际任务 | 代码补全 | 163 |
| 填空 | 2129 | |
| 代码修复 | 76 | |
| 缺陷检测 | 6006 |
排行榜
闭源模型
| 家族 | 模型 | 大小 (B) | MMLU | GSM8k | HumanEval | MBPP | CodeMMLU |
|---|---|---|---|---|---|---|---|
| Anthropic | Claude-3-sonnet@20240229 | - | 88.7 | 96.4 | 92.0 | 76.6 | 55.5 |
| OpenAI | GPT-4o-2024-05-13 | - | 88.7 | 95.8 | 90.2 | 81.4 | 65.0 |
| GPT-3.5-turbo-1106 | - | 61.9 | 73.8 | 61.4 | 78.5 | 51.6 |
开源模型
| 家族 | 模型 | 大小 (B) | MMLU | GSM8k | HumanEval | MBPP | CodeMMLU |
|---|---|---|---|---|---|---|---|
| MetaLlama | Meta-Llama-3,1-70B-Instruct | 70.0 | 83.6 | 95.1 | 80.5 | 75.4 | 59.7 |
| Meta-Llama-3,1-70B | 70.0 | 79.3 | 83.7 | 58.5 | 66.2 | 40.5 | |
| Meta-Llama-3-70B | 70.0 | 79.5 | 83.0 | 48.2 | 70.4 | 49.7 | |
| Meta-Llama-3-70B-Instruct | 70.0 | 82.0 | 93.0 | 81.7 | 82.3 | 61.8 | |
| CodeLlama-34b-Instruct-hf | 34.0 | - | - | 41.5 | 57.0 | 39.3 | |
| Mistral | Mistral-7B-Instruct-v0.3 | 7.0 | 62.5 | 50.0 | 26.2 | 50.2 | 44.1 |
| Mixtral-8x7B-Instruct-v0.1 | 46.7 | 70.6 | 74.4 | 40.2 | 60.7 | 42.7 | |
| Codestral-22B-v0.1 | 22.0 | - | - | 81.1 | 78.2 | 47.6 | |
| Phi | Phi-3-medium-128k-instruct | 14.0 | 78.0 | 91.0 | 62.2 | 75.2 | 48.7 |
| Phi-3-mini-128k-instruct | 3.8 | 68.8 | 82.5 | 58.5 | 70.0 | 39.2 | |
| Qwen | Qwen2-7B-Instruct | 7.0 | 70.5 | 82.3 | 79.9 | - | 51.9 |
| Qwen2-57B-A14B-Instruct | 57.0 | 76.5 | 80.7 | 53.0 | 71.9 | 47.3 | |
| CodeQwen1.5-7B-Chat | 7.0 | - | - | 83.5 | 77.7 | 47.7 | |
| Yi | Yi-1.5-34B-Chat | 34.0 | 67.6 | 71.7 | - | - | 50.0 |
| Yi-1.5-9B-Chat | 9.0 | 68.4 | 52.3 | 39.0 | 54.4 | 48.2 | |
| DeepSeek | DeepSeek-coder-7B-instruct-v1.5 | 7.0 | 49.2 | 41.0 | 42.1 | 60.7 | 41.6 |
| DeepSeek-coder-33B-instruct | 33.0 | - | 60.7 | 79.3 | 70.0 | 37.5 | |
| DeepSeek-moe-16B-chat | 16.4 | 45.0 | 18.8 | 26.8 | 39.2 | 31.5 | |
| DeepSeek-Coder-V2-Lite-Instruct | 16.0 | 60.1 | 86.4 | 81.1 | - | 47.1 | |
| InternLM | InternLM2.5-20B-chat | 20.0 | 66.5 | 79.6 | 48.8 | 63.0 | 46.2 |
| StarCoder | StarCoder2-15B-instruct-v0.1 | 15.0 | - | - | 46.3 | 66.2 | 47.8 |
许可证
该项目基于 MIT 许可证。
引用
如果此仓库对您有帮助,请考虑引用我们的论文:
@article{nguyen2024codemmlu, title={CodeMMLU: A Multi-Task Benchmark for Assessing Code Understanding Capabilities}, author={Nguyen, Dung Manh and Phan, Thang Chau and Le, Nam Hai and Doan, Thong T. and Nguyen, Nam V. and Pham, Quang and Bui, Nghi D. Q.}, journal={arXiv preprint}, year={2024} }

- 1CodeMMLU: A Multi-Task Benchmark for Assessing Code Understanding Capabilities of CodeLLMsFPT软件AI中心,越南 · 2024年



