Layer selection

Overview#

Layer selection names which transformer blocks Ster reads from or trains on. Layers are 0-based decoder-block indices, 0 through num_hidden_layers − 1; the count comes from the loaded model's config.json, so the same selection string can be valid for one model and refused for another.

Grammar#

--layers (on train, optimize, extract; default all) accepts (parse_layers in src/workflow.rs):

Segments are trimmed, empty segments ignored, and the final list is sorted and deduplicated — 5,3,3..6 selects 3, 4, 5. Ranges are half-open, matching Rust: 1..3 is layers 1 and 2, and the README's 12..20 is layers 12–19.

Refusals#

Parsing and validation fail closed, in this order:

• unparsable segment: invalid layer "x" or invalid layer range "8..x";
• reversed or empty range: layer range "8..8" must have start < end;
• nothing selected (e.g. --layers " , "): no layers selected;
• out of range for the loaded model: layer 4 is outside the model's 0..3 range;
• a model reporting zero layers: model has no layers.

The runtime applies the same bound check wherever layer indices arrive from an artifact instead of a flag (validate_layers in src/runtime.rs, plus at least one layer is required), so a vector trained on a deeper model is refused at generation time rather than silently dropped.

Where selection matters#

• `train` writes one direction per selected layer into a single artifact; wide selections make bigger artifacts (each vector is hidden_size floats) and cost one capture per layer per prompt.
• `optimize` treats the selection as the search space and keeps exactly one winning layer, after publishing what every layer in the space scored on the holdout.
• `extract` exports activations for every selected layer.
• `evaluate` and `generate` take no --layers flag: the artifact's own vectors decide which layers are scored or steered.

Which layers encode a trait is empirical — middle layers are a common starting point, but ster vector optimize --layers all is the grounded way to find out, as executed in walkthrough-steering.