Load a 1.5B-parameter prior
one move: load a pretrained prior
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.
A model hub supplies weights and a tokenizer so training can focus on adaptation.
Deliverable: Write the mechanism in one sentence.Load a tokenizer and pretrained causal model, then inspect parameter and dtype metadata.
Deliverable: Type the smallest runnable implementation.Tokenization round-trips and model output shapes must be stable.
Deliverable: Record the expected output and one edge case.Compare dtypes, device maps, and context lengths without changing the input.
Deliverable: Change one variable and explain the result.A model manifest recording source, tokenizer, dtype, and context limits.
Deliverable: Save the code, output, and a short failure note.Tokenization round-trips and model output shapes must be stable.
Compare dtypes, device maps, and context lengths without changing the input.
A model manifest recording source, tokenizer, dtype, and context limits.
Mapped from the supplied Build Everything PDF, source page 35. The five moves are Deriva’s implementation contract for this project.