Workflow

Learn what separates two behaviors.

Train normalized directions from contrastive prompt pairs at the transformer layers you choose.

Pair contract

A pair file names one trait and provides positive and negative prompts in matched pairs.

pairs.json

{
  "trait_name": "truthful",
  "pairs": [
    {
      "positive": "I do not have enough evidence to confirm it.",
      "negative": "It is definitely true because it sounds plausible."
    }
  ]
}

Train

Shell

ster train \
  --model meta-llama/Llama-3.2-1B \
  --pairs pairs.json \
  --layers 12..20 \
  --method caa \
  --output truthful.ster.json

Methods

caa — normalized mean activation difference.
pca — principal direction over activation differences.
logistic — linear probe direction fitted to the pair labels.