[EDIT] Unfortunately authors have withdrawn the publication due to data leakage. The original article follows.
NewHope, a chat model based on Llama-13b, developed by the SLAM research group at Shanghai University of Finance and Economics, demonstrates 99% of GPT-4’s programming capabilities. The model can handle multiple programming languages and has been open-sourced for public use. Preliminary evaluations on HumanEval show promising results, with NewHope scoring 66.5 on the Pass@1 metric, closely following GPT-4’s score of 67.0.
Read more at GitHub…