Adapt a model with a tiny trainable footprint
one move: adapt without rewriting
Open the code pane immediately, then use the Spec tab, optional design question, tests, and artifact when you want them.
Finish each move before carrying the artifact forward.
LoRA freezes base weights and learns a low-rank update W′=W+BA.
Deliverable: Write the mechanism in one sentence.Insert low-rank adapters, freeze the base model, and fine-tune on a small task set.
Deliverable: Type the smallest runnable implementation.Only adapter parameters should change and validation must be reproducible.
Deliverable: Record the expected output and one edge case.Vary rank, adapter placement, and learning rate; compare quality versus trainable bytes.
Deliverable: Change one variable and explain the result.An adapter checkpoint, parameter budget, and before/after evaluation.
Deliverable: Save the code, output, and a short failure note.Only adapter parameters should change and validation must be reproducible.
Vary rank, adapter placement, and learning rate; compare quality versus trainable bytes.
An adapter checkpoint, parameter budget, and before/after evaluation.
Mapped from the supplied Build Everything PDF, source page 50. The five moves are Deriva’s implementation contract for this project.