【北京】北京大学的aiXcoder团队近日推出了一项重大技术创新——aiXcoder-7B Base版代码大模型，这是一款专为企业私有部署设计的开源AI解决方案。该模型在海量的1.2T Unique Tokens数据集上进行了深度训练，确保了其在实际软件开发场景中的高效应用。设计上，aiXcoder-7B Base充分考虑了真实代码生成的需求，其预训练任务和上下文信息都经过精心定制，以满足企业级的复杂要求。
在性能评测方面，aiXcoder-7B Base的表现令人瞩目。在HumanEval、MBPP和MultiPL-E三大业内公认的代码生成评测集上，该模型的平均得分超越了拥有340亿参数的知名模型Codellama。特别是在代码补全任务上，aiXcoder-7B Base在同等参数量的模型中脱颖而出，展现出卓越的性能。
此外，aiXcoder-7B Base在多语言自然语言到代码转换（NL2Code）基准测试中的平均效果也超越了Codellama 34B和StarCoder2 15B。这一成就标志着北大aiXcoder团队在AI代码生成领域的领先地位，也为企业级代码智能辅助开发提供了全新的选择。
北京大学的这一开源项目，不仅展示了中国在人工智能领域的科研实力，也为全球企业提供了更加先进、安全和定制化的代码开发工具，有望推动软件开发行业的效率与质量提升。
英语如下：
**News Title:** “Peking University’s Open-Source Coding Master: aiXcoder-7B, A New Benchmark for Enterprise Code Generation”
**Keywords:** aiXcoder-7B, Enterprise Deployment, Code Generation
**News Content:** **Beijing** – The aiXcoder team from Peking University recently unveiled a major technological innovation, the aiXcoder-7B Base Edition, a large language model specifically designed for enterprise-scale private deployment. This open-source AI solution has been deeply trained on a massive 1.2T Unique Tokens dataset, ensuring its effectiveness in real-world software development scenarios.
In its design, the aiXcoder-7B Base takes into account the demands of actual code generation, with its pre-training tasks and contextual information meticulously tailored to cater to the complexities of enterprise-level requirements.
In terms of performance, the aiXcoder-7B Base has made a significant impact. Outperforming the well-known 340 billion parameter model, Codellama, on the industry-acknowledged code generation benchmark tests HumanEval, MBPP, and MultiPL-E, this model averages higher scores.特别是在代码补全任务中，aiXcoder-7B Base, with comparable parameter count, demonstrates exceptional performance.
Furthermore, the aiXcoder-7B Base surpasses Codellama 34B and StarCoder2 15B in the average performance across multiple language natural language to code (NL2Code) benchmarks. This milestone underscores Peking University’s aiXcoder team’s leading position in the AI code generation field and presents a new option for enterprise-level intelligent code-assisted development.
Peking University’s open-source endeavor not only exhibits China’s research prowess in artificial intelligence but also offers global enterprises more advanced, secure, and customizable code development tools. This innovation has the potential to enhance the efficiency and quality of the software development industry.
【来源】https://www.qbitai.com/2024/04/134070.html
Views: 1