“However, LLaMA and all its derivatives (including Alpaca, Vicuna, and Koala) are only available for non-commercial research purposes.”
The LLaMA paper describes its training data in public, while the models stay research-only. Together, Ontocord.ai, ETH DS3Lab, Stanford CRFM and Hazy Research followed that recipe and rebuilt the 1.2 trillion token dataset, about 5TB unzipped, for anyone to use. Every Alpaca and Vicuna built on LLaMA is stuck in the lab. A license fence only holds until five organizations decide to rebuild the data from CommonCrawl.