Distillation
Named by 17 kept items in this workspace.
The items that name it 17
Training a student model on responses collected from a teacher model through its API, the copying scenario WM-Basin targets.
Generating millions of responses from a larger model and using them to teach a smaller model to mimic its behavior. Illicit distillation does this without the developer's authorization, often via networks of fraudulent accounts.
Harvesting model outputs, including chain-of-thought, to train other models, as Anthropic reports by PRC labs using proxy farms.
via jailbroken APIs
In every workspace you can read 18
AI
All techniques in this workspace.