← back to test matrix

Test 0.7

Can we get free accuracy by averaging three of our model checkpoints?

No. The soup over-detects badly — precision falls to about 50% versus 83-88% for any single sibling model. It draws roughly 4 extra junk lines per photo. Not free accuracy; it's a worse model.
What it means for SailScan

Weight-averaging our three sibling checkpoints sounded like a zero-cost accuracy bump — same serving setup, no retraining. It isn't: the averaged model gets confused and draws far too many candidate lines, most of them wrong. This path is closed; don't spend more time on model soups from these checkpoints.

49.7%
soup precision
83.2%-87.7%
sibling precision range
NO
verdict
The detail
ModelRecallPrecisionAvg line error
v1199.8%85.5%0.0072
v15 (best single sibling)99.9%83.2%0.0063
v2099.8%87.7%0.0069
Soup (v11+v15+v20 averaged)100.0%49.7%0.0105

A "model soup" averages the weights of several sibling models into one, hoping to get the best of all three for free. Here it matched or slightly beat recall, but precision collapsed — the soup draws roughly twice as many false lines as any single sibling.

Source: 332-photo holdout, frozen 2026-08-25. Raw model output, no production rules filter applied.

raw data (JSON) →