【北京讯】北京大学的aiXcoder团队近日宣布，他们成功开源了全新的aiXcoder-7B Base版代码大模型，这款模型专为企业私有部署而设计，旨在提升企业软件开发的效率与质量。aiXcoder-7B在1.2T Unique Tokens的海量数据上进行了深度训练，其训练过程与预训练任务都针对真实的代码生成场景进行了定制，以满足企业级应用的需求。
据量子位报道，aiXcoder-7B在HumanEval、MBPP和MultiPL-E三大主流代码生成评测集上的表现优于拥有340亿参数的Codellama。在代码补全这一关键功能上，aiXcoder-7B展现出了同等级参数量模型中的最佳效果。此外，在多语言自然语言到代码转换（NL2Code）基准测试中，aiXcoder-7B的平均性能也超越了Codellama 34B和StarCoder2 15B，彰显了其在跨语言编程任务中的强大实力。
这一创新成果的发布，标志着我国在人工智能辅助软件开发领域的又一重大突破，有望为企业提供更加智能化、高效化的代码生成解决方案，进一步推动我国乃至全球的软件开发行业向前发展。北大aiXcoder团队的这一开源举措，也展现了学术界与产业界合作推动技术创新的积极态势。
英语如下：
**News Title:** “Peking U’s aiXcoder-7B: A New Benchmark in Enterprise Code Generation Models, Breaking Records”
**Keywords:** Peking U’s aiXcoder, Large Code Models, Enterprise Private Deployment
**News Content:** **[Beijing]** The aiXcoder team from Peking University recently announced the open-source release of their cutting-edge aiXcoder-7B Base Edition, a large code model specifically designed for enterprise private deployment. Aimed at enhancing efficiency and quality in corporate software development, the model has been deeply trained on a massive dataset of 1.2T Unique Tokens, with both its training process and pre-training tasks tailored for real-world code generation scenarios to meet enterprise-level requirements.
According to QbitAI, aiXcoder-7B outperforms Codellama, a model with 340 billion parameters, on the three mainstream code generation benchmarks—HumanEval, MBPP, and MultiPL-E. Demonstrating superior performance in code completion, aiXcoder-7B achieves the best results among models with comparable parameter counts. Furthermore, in the multilingual Natural Language to Code (NL2Code) benchmark tests, aiXcoder-7B surpasses both Codellama 34B and StarCoder2 15B, highlighting its prowess in cross-language programming tasks.
This innovative breakthrough signifies another significant advancement in China’s AI-assisted software development domain and is poised to offer enterprises more intelligent and efficient code generation solutions. It is expected to further propel the software development industry both domestically and globally. The open-source initiative by the Peking U’s aiXcoder team exemplifies the positive momentum of collaboration between academia and industry in driving technological innovation.
【来源】https://www.qbitai.com/2024/04/134070.html
Views: 1