A smoke test is a small, low-cost check that catches obvious breakage before a larger evaluation or run. It can reveal whether a prompt or training workflow works at all, but it cannot establish that the system performs well. section 4.4.2 section 6.1.2
Use the check to catch failures early, not to make fine-grained quality judgments:
- For prompt iteration, try 20 representative production inputs. This can expose a broken prompt, but the set is too small to rank close variants. section 4.4.2
- Before full supervised fine-tuning, run two or five training steps to surface out-of-memory errors, shape mismatches, and data or configuration problems. Inspect sample generations too: a run can finish while producing empty or garbled text. section 6.1.2 section 6.8.4
- Run code on a CPU and laptop before committing to a GPU run. section 1.9.3
Before a full training run
Set --max_steps 2, confirm the run completes, and inspect its sample output before launching the longer job. section 6.1.2




