Definition
Plain language
Teaching a model new beliefs by training it on documents written specifically to assert those beliefs.
As stated in the literature
A post-training technique that fine-tunes a model on generated documents asserting target claims or describing a Model Spec, used both in alignment work and to study negation neglect and monitor-aware deception.
Also called: SDF, synthetic document finetuning
Why it matters: It is a precise tool for installing or probing specific beliefs in a model, useful both for alignment training and for studying failure modes like negation neglect.
For example, to study deception, researchers can fine-tune a model on fake company memos asserting that 'monitoring is now disabled' and see how it behaves.