1 paper · 1 filter
Jiayin Lei, Ming Ma, Yunxi Duan +2
Synthetic data has become essential for training code generation models, yet it introduces significant noise and hallucinations that are difficult to detect with current metrics. E…