Yongcheng Huang
.png)
.png)
This study evaluates code LLMs for multilingual comment generation, reveals evaluation biases in non-English languages, and offers an open-source dataset for improvement.
This study evaluates the performance of LLMs in generating Chinese Java code comments, introduces a detailed error taxonomy, and demonstrates that cosine similarity of word embeddings can help automatically detect semantic errors like hallucinations.
In this paper, we propose an intelligent trigger model that combines rule-based filtering with deep learning on telemetry and code context data to efficiently decide when to invoke a large language model for code completion, reducing unnecessary computations while enhancing developer productivity.
The Dfusion structure is modeled after the Mixture of Experts concept of LLM, which integrates the results of learning multiple SOTA models and achieves results that surpass the current SOTA through extremely lightweight CNN and extremely low training load.
Projects
Alpha-built is designed to be next generation of online, multi-worker, light-weight architecture design platform. With git like version control system, we are able to empower the future of architecture design.
During this challenge, we implemented a method of implementing segmentation tasks in Jetson Nano with only 5w power consumption.
Green Llama is a tool that monitors the energy and environmental impact of LLMs—tracking CPU, GPU, and RAM usage—to help optimize efficiency and reduce carbon emissions.
Achievements
