Train a LLM from Scratch

Curious — how did you handle training stability early on? Was convergence an issue without heavy tuning?