Liyang Fan

Education

• Specialized in Computer Technology
• Research conducted at Shenzhen Institute of Advanced Technology

• Ranked in the top 1% of the program
• Key courses: Software Engineering, Discrete Mathematics, Data Structures, Algorithm Design

Work Experience

• Worked on model architecture adaptation, data mining, and feature extension
• Contributed to the development of jina-reranker-v2 series, available at https://huggingface.co/collections/jinaai/jina-reranker-v2-667a7e6b8a3c33a936e9c99b
• Achieved significant progress especially in code, tool, and table-related tasks
• Key achievements include:
• CodeSearchNet (MRR@10, 3 tasks)
• ToolBench (recall@3, 3 tasks)
• TableSearch (recall@3)
• Contributed to the development of jinaai/reader-lm series, which is a small model for converting HTML to Markdown, bringing greater convenience for data acquisition. Available at https://huggingface.co/collections/jinaai/jina-reader-lm-670628bfe0d782685cb53416

Projects

AgentCourt

• Designed and implemented the overall front-end and back-end architecture for a multi-agent system
• Developed inter-agent communication protocols and dialogue systems
• Built and integrated knowledge base systems for enhanced agent interactions

Multi-GPU PDF Processing Pipeline

• Developed an end-to-end PDF processing system using multi-process and multi-GPU architecture
• Built on the Mineru framework to optimize document processing efficiency
• Implemented parallel processing techniques for improved performance

ChatAgent

• Developed an AI agent project based on FastAPI framework
• Implemented natural language database querying, improving data retrieval efficiency
• Achieved response times under 3 seconds for queries on tables with millions of records

MegaChat Text Generation Model

• Built a text generation system based on the MEGABYTE model
• Implemented efficient inference with extended length capabilities
• Developed character-level inference for multi-language support

ByteLLM

• Developed ByteLLM, focusing on end-to-end training of Large Language Models using ByteTokenizer
• Integrated byte-level tokenization into LLMs to simplify and enhance the training process
• To Demonstrate compatibility with existing LLM frameworks, facilitating seamless integration
• To Achieve significant improvements in tokenization and training efficiency

Skills

• Programming Languages: Python, C++, Java
• Machine Learning: Experienced in designing and implementing ML models
• Frameworks: PyTorch, FastAPI, Hugging Face Transformers
• Tools: Git, Docker

Publications

Liu, Z., Fang, F., Feng, X., Du, X., Zhang, C., Wang, Z., Bai, Y., Zhao, Q., Fan, L., Gan, C., et al. (2024). II-Bench: An Image Implication Understanding Benchmark for Multimodal Large Language Models. arXiv preprint arXiv:2406.05862.