loss drop to zero when training background

[nothonfz]

When I use custom data for the first stage of training, losses like coarse_empty_space_loss become zero at around 60,000 steps, and the rendering results for validation are also extraordinary. Do you know if this is normal? What could be causing this?