Ster
Ster/Docs

Workflow

Select the method and layer on held-out pairs.

Ster ranks candidate directions by pair-ordering accuracy, then projection margin, before retraining the winner on every pair.

Selection

Optimization requires at least four pairs and uses an 80/20 split. Every requested layer is compared across CAA, PCA, and logistic training.

Shell
ster optimize \
  --model meta-llama/Llama-3.2-1B \
  --pairs pairs.json \
  --layers all \
  --output best.ster.json

Ranking invariant

  • Higher holdout pair-ordering accuracy wins first.
  • Projection margin breaks equal-accuracy candidates.
  • The selected method and layer are retrained on the complete pair set.