Layer and width must stay explicit
Ster reads last-token hidden states from selected transformer layers and records hidden width in the artifact.
Ster by Wisent
Ster is a native Rust toolkit for reading last-token hidden states, learning directions from contrastive examples, evaluating them, and applying measured interventions during generation.
Representation work breaks when model, layer, width, method, and intervention point drift apart.
Ster reads last-token hidden states from selected transformer layers and records hidden width in the artifact.
Artifacts record trait, method, layer-indexed directions, training accuracy, and projection margin.
Ster refuses artifacts with the wrong product, schema, model, or vector width instead of applying them.
The workflow stays inside one native Rust binary and library crate with versioned JSON artifacts.
Ster reads last-token hidden states from selected transformer layers.
Ster learns directions from contrastive examples with CAA, PCA, or logistic-probe training.
Ster evaluates artifacts with pair-ordering accuracy and projection margin, with holdout selection across method and layer.
Ster adds a measured direction in the residual stream during autoregressive generation.
One binary and one library crate share the same versioned JSON artifacts and native Candle runtime.
| Operation | Role in workflow |
|---|---|
| train | Learn directions from examples |
| optimize | Select across method and layer |
| evaluate | Score ordering and margin |
| generate | Apply steering during decoding |
| extract | Read selected hidden states |
| inspect | Read recorded fields |
Every steering artifact records identity and measurement fields before generation uses it.
| Artifact field | Boundary |
|---|---|
| Product identity | Wrong product refused |
| Schema identity | Wrong schema refused |
| Model revision | Wrong model refused |
| Hidden width | Wrong width refused |
| Method and trait | Kept with directions |
| Accuracy and margin | Recorded before steering |
No. Ster controls local open-weight models; hosted routing, fleet placement, credentials, and release delivery belong to other products.
The runtime supports local and Hugging Face Llama-family Safetensors checkpoints on CPU, with compile-time Metal and CUDA backends.
Ster refuses artifacts with the wrong product, schema, model, or vector width instead of applying them to the wrong residual stream.
Use Ster when you need a native Rust path from hidden states and contrastive examples to measured artifacts and controlled generation, with refusal at the artifact boundary.