Is anyone else getting this kind of corruption on on-site training?

Tried this dataset (with a few tweaks) on Illustrious twice. First time I got completely corrupted results for one sample prompt (and verified even past sample images one of the tags in that sample consistently produced this corrupt outfit with the LoRA loaded). Second time every single sample picture past first couple epoches was completely corrupted. Is this a widespread issue?

This is proving quite annoying and demoralizing to me because the thing I do when annoyed and demoralized is work on my hobbies, and AI Toolkit isn't letting me do that (and I will freely blame because I have never had this issue with Kohya in all my training runs on it. Single random epoches failed to render, but never back to back to back like seen here)