Yi-Lightning Technical Report
01. AI, Alan Wake, Albert Wang, Bei Chen, C. X. Lv, Chao Li, Chengen Huang, Chenglin Cai, Chujie Zheng, Daniel Cooper, Ethan Dai, Fan Zhou, Feng Hu, Heng Ji, Howard Qiu, Jiangcheng Zhu, Jun Tian, Katherine Su, Lihuan Zhang, Liying Li, Ming Song, Mou Li
2024-12-02
Summary
This paper presents Yi-Lightning, a new large language model (LLM) that performs exceptionally well in various categories, including Chinese language tasks, math, coding, and complex prompts.
What's the problem?
While many large language models exist, they often struggle with specialized tasks and may not perform consistently across different subjects. Additionally, training these models can be expensive and time-consuming, which limits their practical use in real-world applications.
What's the solution?
Yi-Lightning addresses these issues by using an advanced architecture called Mixture-of-Experts (MoE), which allows the model to use different 'experts' for different tasks without needing excessive resources. The development process includes comprehensive pre-training, supervised fine-tuning, and reinforcement learning from human feedback. The model also incorporates a safety framework called RAISE to ensure responsible AI use. These innovations help reduce costs while maintaining high performance.
Why it matters?
This research is important because it demonstrates how to create more effective and efficient AI models that can handle a wide range of tasks. By improving the performance of LLMs like Yi-Lightning, we can enhance applications in education, healthcare, and many other fields where accurate language understanding is crucial.
Abstract
This technical report presents Yi-Lightning, our latest flagship large language model (LLM). It achieves exceptional performance, ranking 6th overall on Chatbot Arena, with particularly strong results (2nd to 4th place) in specialized categories including Chinese, Math, Coding, and Hard Prompts. Yi-Lightning leverages an enhanced Mixture-of-Experts (MoE) architecture, featuring advanced expert segmentation and routing mechanisms coupled with optimized KV-caching techniques. Our development process encompasses comprehensive pre-training, supervised fine-tuning (SFT), and reinforcement learning from human feedback (RLHF), where we devise deliberate strategies for multi-stage training, synthetic data construction, and reward modeling. Furthermore, we implement RAISE (Responsible AI Safety Engine), a four-component framework to address safety issues across pre-training, post-training, and serving phases. Empowered by our scalable super-computing infrastructure, all these innovations substantially reduce training, deployment and inference costs while maintaining high-performance standards. With further evaluations on public academic benchmarks, Yi-Lightning demonstrates competitive performance against top-tier LLMs, while we observe a notable disparity between traditional, static benchmark results and real-world, dynamic human preferences. This observation prompts a critical reassessment of conventional benchmarks' utility in guiding the development of more intelligent and powerful AI systems for practical applications. Yi-Lightning is now available through our developer platform at https://platform.lingyiwanwu.com.