Chantal Pellegrini

For me, batch evaluation also gave nan values, after fine-tuning. Performing evaluation also with bfloat16 instead of float16 solved this for me. (I also fine-tuned using bf16 True). Like this,...