Ster
Ster/Docs

Direction (steering vector)

A direction is a unit-length f32 vector in activation space — one per layer — that points from "negative" toward "positive" for a trait. It is Ster's core object: reading projects onto it, steering adds along it. In artifacts it appears as

Overview#

A direction is a unit-length f32 vector in activation space — one per layer — that points from "negative" toward "positive" for a trait. It is Ster's core object: reading projects onto it, steering adds along it. In artifacts it appears as a LayerVector; the README and CLI call the whole per-layer collection a steering vector.

Shape#

Inside a steering artifact (LayerVector in src/artifact.rs):

json
{
  "layer": 3,
  "values": [0.10141639, 0.16717353, -0.03978117, ...],
  "train_margin": 0.92032135,
  "train_accuracy": 0.875
}
  • layer — the 0-based block index the vector was read from and will be added to.
  • values — exactly hidden_size finite floats.
  • train_accuracy, train_margin — the evaluation scores measured at save time, kept with the vector so an artifact is self-describing.

Lifecycle#

  1. Trained — train_direction (src/representation.rs) turns matched positive/negative activation rows into a raw direction using one training method.
  2. Normalized — every method's output is L2-normalized before it leaves training. This is unconditional; a direction whose norm is zero or non-finite is refused (training produced a zero or non-finite direction). Because directions are unit-length, --strength is the entire magnitude knob and strengths are comparable across methods.
  3. Scored and saved — accuracy/margin are computed on the training pairs and stored beside the values.
  4. Applied — at generation, the vector is scaled by strength and added to the residual stream at its layer (intervention).

Sign convention#

Positive projection = more of the trait, because training subtracts negative from positive (caa averages pos − neg; pca sign-aligns its principal component with that mean difference; logistic labels positives 1). So --strength 1.0 pushes generation toward the trait and --strength=-1.0 pushes away. Keep the convention in the pair file, not in your head: if the columns are swapped, the direction is simply negated.

Invariants and refusals#

Enforced by artifact validation and the runtime gate:

  • Width must match the artifact's hidden_size: layer 1 has vector width 10, expected 64.
  • Every component finite: layer 3 contains a non-finite value.
  • Width must match the loaded model at apply time: artifact width 64 does not match model width 2048 and, at plan construction, layer 3 steering vector width 10 does not match model width 64.
  • The layer must exist in the loaded model: layer 4 is outside the model's 0..3 range.

Directions are comparable with ster::representation::cosine_similarity (library API): it refuses mismatched or empty inputs (cosine similarity requires equal non-empty vectors) and zero vectors (cosine similarity is undefined for a zero vector).