fine tuning

Three ways to give a model your data, three different places the data ends up. The answer that scores follows the data and asks how you would delete it.

A training update can change behavior beyond its intended task. Work through a release comparison that separates useful answers, forbidden proposals and confirmed effects.

Fine-tuning moves data from a database you can query and delete into weights you cannot. The scored answer covers memorisation, tenant separation, poisoning and the deletion problem.

The model passes your evaluation because your evaluation does not contain the trigger. And safety training may teach it to hide better.

'Clean the dataset' is not an investigation. The answer freezes evidence, traces what changed and who could change it, tests causality by retraining, and treats the pipeline as possibly compromised too.