Hugging Face rejected my 2nd-order optimizer PR for being "too new", so I made it a 1-line standalone drop-in for local LoRA fine-tuning

Recently, I developed SCAO (Sparse Curvature-Aware Optimizer), a 2nd-order optimizer designed to fix the slow early-stage convergence of AdamW when fine-tuning LLMs. I tried to get it integrated into transformers, but the maintainers understandably rejected the PR. The feedback was essentially: "It’s too new, the math is complex, and we need to see concrete community adoption before adding it to the core library." Fair enough. So I removed the friction and made it a standalone script. If you