A guy in a Discord chat said to freeze my training data before fine-tuning, and he was right
Back in March I was fine-tuning a small text model for my side project and someone in a Discord chat named Dev_Ari said "always snapshot your dataset first, you'll thank me at 2am" so I did. Two runs later my labels got scrambled and I had a clean copy to fall back on, which saved me about 6 hours of relabeling. Has anyone else had a random stranger's one-liner save their whole project like that?