ChatGLM-6B full parameter fine-tuning loss = 0

[xiaojunjun65]

Today, when I was doing the full parameter fine-tuning, my loss became 0.0 after the training of batch 17, and then it kept going to 0.0. I changed both the learning rate and batch, but I still failed to solve such problems

[shibing624]

try to set float32, not fp16. i run full parameter fine-tuning success with llama, not chatglm ever.