Teaching a model what good looks like

Supervised fine-tuning trains a model on examples of desired responses. For an LLM, those examples can include instructions, tool calls, worked solutions, and complete trajectories. The target is the demonstrated behavior, so the data distribution and its blind spots shape what the model learns.

Decide which failure you are trying to repair. A model that cannot produce a valid tool call needs a different dataset from one that produces valid calls but chooses the wrong action. Preserve task context, tool observations, and the reason an action is appropriate. Deduplicate at the problem and source level before separating training from evaluation.

Evaluate the exact base checkpoint and the tuned checkpoint on the same held-out items under matched inference settings. Track completion reliability as well as substantive correctness. Better final-answer completion can raise a headline score without an equally large gain in underlying reasoning.

Start with a held-out failure taxonomy, train on targeted demonstrations, and compare against a similarly sized general-data intervention. Keep token budget and evaluation contract fixed. Report whether gains transfer across task families and whether previously reliable behaviors regress.