From what I can tell they're fine-tuning EleutherAI's GPT-J. Alpaca was made to fine-tune LLaMa, however they also released their dataset they used to do this,

From what I can tell they're fine-tuning EleutherAI's GPT-J. Alpaca was made to fine-tune LLaMa, however they also released their dataset they used to do this, and it looks like Dolly is this dataset applied to GPT-J, and does not use LLaMa itself.