A "work-in-progress" collection of experimental SFT runs. Primary focus: minimizing catastrophic forgetting, testing LoRA vs. Full-Parameter tuning.