tatsu-lab/stanford_alpaca ? reverse-engineered prompt

Build me a working repo for Stanford Alpaca that can both generate the instruction tuning data and fine tune an instruction following LLaMA style model on it.

I want the project to let me create the 52K style dataset from seed tasks using the OpenAI API key, save it as JSON, then run a training script that fine tunes a model with the provided Alpaca prompts and settings. If possible, also include the option to recover the Alpaca weights from a released weight diff.

Keep it research focused and make the scripts easy to run from the command line, with clear defaults, helpful error messages, and a simple README that explains the full flow from data generation to training. If you need to check current library docs online, go ahead.