Duration
Start time: May 2024 End time: August 2024
Projects Followed:
Experience at Digital China
● Developed full function calling finetuning data generation pipeline based on 14 corporate APIs, and successfully generated dataset with 20,000 more usable data for fintuning. ● Finetuned Qwen2-7B-agent-Instruct prompt version and GLM-4-9B-chat FC version with Llama Factory using synthetic data(generated with pipeline inspired by APIGen), improved model’s function calling ability from 86.42 to 92.14(Qwen2-7b-instruct), and 10.7 to 81.24(Qwen2-1.5B-Instruct) , score based on Berkley Function Calling Leaderboard. ● Edit and modify BFCL to enable local LLM inference and evaluation, modify the framework and deployed to company product ShenZhouWenXue. ● Continually follow up on AI community products and AIGC products, and conduct experience testing.