Ster/Docs

Product contract

Control the residual stream, deliberately.

Ster is a native Rust toolkit for reading hidden states, learning representation directions, measuring them, and applying them during generation.

What Ster owns

Ster owns the local representation-engineering loop for open-weight language models. It loads a checkpoint, captures final-token residual states after selected transformer blocks, learns layer-indexed directions, and applies a selected direction before the next block.

The runtime uses Candle directly so tokenization, attention, KV caching, sampling, activation capture, and steering remain in one native process.

Current surface

  • Llama-family Safetensors checkpoints from a local directory or Hugging Face Hub.
  • CAA, PCA, and logistic-probe direction training.
  • Holdout selection across method and layer.
  • Pair-ordering accuracy and projection-margin evaluation.
  • Additive residual-stream steering during autoregressive generation.
  • Versioned JSON pair, activation, steering, and evaluation artifacts.

Boundaries

  • Unsupported model architectures fail before weights are loaded; Ster does not guess adapters from tensor names.
  • Hosted model routing belongs to Brama.
  • Fleet placement and job lifecycle belong to Stado.
  • Credentials belong to Skarbiec.
  • The retired Python package and wisent command are not compatibility surfaces.