TL;DR: Stick with PEFT. Full fine-tuning is risky for deadlines. I suggest using NVIDIA A100 80GB GPU for your QLoRA runs, just be careful with overfitting your legal data.

Honestly, skip the full fine-tuning. For legal docs, you dont need the massive overhead of full parameters unless you are doing weird language translation stuff. I have spent years training models on heavy technical jargon and QLoRA is honestly way more than enough if you prep your dataset right. Its just way faster and keeps costs down, which sounds like exactly what you need for a deadline next week. QLoRA is the move because it saves your budget while keeping accuracy high. You can run this on a solid GPU like the NVIDIA A100 80GB PCIe without hitting your bank account too hard. Full fine-tuning is just a pain in the neck unless you have a massive cluster sitting around and months to waste on hyperparameter tuning. If you really want that extra edge, look into LoRA+ or even just standard LoRA with higher rank settings for the specific legal layers. In my experience, the secret sauce isnt the method, its the data cleaning. Spend those extra hours scrubbing your legal documents for duplicates and formatting errors instead of burning credits on a full run. You will see better results, trust me. If you get stuck on the training loop, just ping me here. You have got this.

Commenting to find later