Huan Song
Kwai Large Model Algorithm Specialist
He graduated from Chinese Academy of Sciences with a master's degree in 2018, and has been engaged in research in the direction of code intelligence since graduation, focusing on code model training and application landing. He joined Racer in 2024, and is responsible for the algorithm research and development of KwaiCoder big model Pre-training, Post-training and Code Agent, etc. At present, his self-developed code big model is better than the open source SOTA base model. Recently, he focuses on the exploration and implementation of cutting-edge LLM algorithms, and constantly breaks through the performance bottleneck of the code base model.
Topic
KwaiCoder Model: A Practical Exploration of Building Advanced Code Macromodels at Low Costs
Achieving current state-of-the-art (SOTA) level performance in code-related tasks, traditional approaches consume significant resources, including large datasets, significant computational power, and complex training processes. We would like to give an overview of our hands-on exploration of using low-cost building of code macromodels, and we expect this work to shed some light on the future development of building code macromodels in the industry. Achieving current state-of-the-art (SOTA) level performance in code-related tasks, traditional approaches consume significant resources, including large datasets, significant computational power, and complex training processes. We would like to give an overview of our hands-on exploration of using low-cost building of code macromodels, and we expect this work to shed some light on the future development of building code macromodels in the industry. 1. The traditional From Cratch approach to building code big models and the challenges encountered; 2. The technical evolution path of building SOTA code big model at low cost and the effect of online use in Kwai; 3. The future planning and outlook of SOTA code big model.
