“Stable LM is trained on a new experimental dataset built on The Pile, but three times larger with 1.5 trillion content tokens. We will release details on the dataset in due course.”
Stable LM ships at 3 and 7 billion parameters under CC BY-SA 4.0, with 15 to 65 billion promised later. The open model runs on a dataset nobody is allowed to see yet. The chat-tuned versions are research-only under a noncommercial license because they lean on Alpaca and ShareGPT. Stability sells transparency and keeps the training data in a drawer.