The LEAP fine-tune loop
Task → data → train → eval → deploy: a new specialist head in minutes, gated by evals before it serves.
The recipe
Define the task contract first (input shape, output schema, eval cases). Generate or collect training data against that contract, audit its distribution before any training run, fine-tune with LoRA on the 350M base, and gate deployment on the eval suite — a specialist that fails its regression cases never reaches a registry.
Honest rough edge: small models are prompt-format sensitive. Evaluate through the exact serving path (same prompt template, same server), not a lab harness — format mismatch between train and serve has cost double-digit accuracy in this codebase before it was made a rule.
When to use it
Whenever a bounded task appears more than a few times a day. The loop's economics are the platform's third aha: adaptation in minutes, owned weights at the end.