【华为创新引领，新型大语言模型“盘古-π”超越国际基准】
华为诺亚方舟实验室近日携手业界，共同推出了一项重大技术创新——盘古-π大语言模型架构。这一创新是对传统Transformer架构的深度革新，通过增强非线性处理，有效解决了特征塌陷问题，从而显著提升了模型的输出表达能力。
盘古-π架构的优越性在实际应用中得到了验证。在相同的数据训练条件下，盘古-π（7B参数量）在多任务性能上超越了具有同等规模的LLaMA 2大模型，实现了10%的推理加速，效率提升明显。尤其值得一提的是，盘古-π在1B参数规模上，已经达到了当前的最优性能水平（SOTA），展现出强大的语言理解和生成能力。
此外，华为团队基于盘古-π架构，还专门研发了一个专注于金融法律领域的应用大模型——“云山”。这一模型的诞生，预示着在金融法律等专业领域，人工智能将能提供更为精准和高效的语言服务，有望推动相关行业的智能化进程。
华为此次的技术突破，再次彰显了中国企业在人工智能领域的研发实力和创新能力，同时也为全球大语言模型的发展树立了新的标杆。未来，盘古-π架构及其衍生应用有望在更多场景中发挥重要作用，持续推动人工智能技术的进步。
英语如下：
News Title: “Huawei Launches Pangu-π Architecture: Exceeding LLaMA to Create a Powerful Financial and Legal Large Language Model”
Keywords: Huawei Pangu-π, large language model, performance surpassing
News Content: **Huawei Leads Innovation with Pangu-π, a New Large Language Model Outperforming International Benchmarks**
Recently, Huawei’s Noah’s Ark Lab has collaborated with industry partners to introduce a groundbreaking technological innovation — the Pangu-π large language model architecture. This innovation represents a profound renovation of the traditional Transformer architecture, enhancing non-linear processing to effectively address the issue of feature collapse, thereby significantly improving the model’s output expression capabilities.
The superiority of the Pangu-π architecture has been demonstrated in practical applications. Under identical data training conditions, Pangu-π (with 7 billion parameters) surpassed the equally scaled LLaMA 2 model in multi-task performance, achieving a 10% inference acceleration, indicating a clear efficiency boost. Notably, Pangu-π at a 1 billion parameter scale has already reached the state-of-the-art (SOTA) performance level, demonstrating its formidable language understanding and generation abilities.
Furthermore, Huawei’s team, leveraging the Pangu-π architecture, has specifically developed a large language model focused on the financial and legal domain named “Yunshan.” This model’s introduction suggests that in specialized areas like finance and law, AI is poised to offer more precise and efficient language services, potentially accelerating the智能化进程 in these industries.
Huawei’s technological breakthrough underscores the company’s research and development strength and innovation capabilities in the field of artificial intelligence. It also sets a new benchmark for global large language model development. In the future, the Pangu-π architecture and its derivative applications are expected to play crucial roles in various scenarios, continually advancing the progression of AI technology.
【来源】https://mp.weixin.qq.com/s/Beg3yNa_dKZKX3Fx1AZqOw
Views: 1