ster vector ablate
Write a steering direction out of a checkpoint's weights, so the model carries the change with no artifact beside it.
Usage#
ster vector ablate --model <MODEL> --vector <VECTOR> --strength <S> --output <DIR> [--revision <REVISION>]What wisent's modify-weights did by directional projection. At every layer the artifact carries, the two projections that write the residual stream — the attention output and the feed-forward down projection — are rewritten as W - strength * v (v^T W) for the layer's unit direction v. At strength one the layer can no longer write along the direction; a smaller strength removes that share of it, and a negative one amplifies it. Ster states no strength. No decoder is built: the tensors are read, rewritten in F32 and cast back to the source dtype, and written as an ordinary checkpoint directory by the same code `ster tune merge` writes with, beside the source's own config, tokenizer and chat template, so every loader reads it.
What it reports#
rewritten— every tensor rewritten, by its stored name.copied_tensors,total_tensors,parameters— what else the checkpoint holds.layers,strength,trait_name,artifact,model,model_revision,output.files— every file written, relative to the output directory.
Desktop request#
ster request vector/ablate takes model, optional revision, vector, strength and output, and returns the same report. Ster Desktop runs it from the Ablate action.
Refusals#
| Sentence | Cause |
|---|---|
ablation strength <S> changes nothing or nothing sensible; state a finite, non-zero strength | A zero or non-finite strength. |
artifact was trained for model <A>, current model is <M> | The artifact belongs to another model; artifact width <W> does not match model width <W> for another width. |
layer <L> is outside a model of <N> layers | The artifact names a layer the model lacks. |
<MODEL> is a GGUF checkpoint; a direction is written into safetensors weights only | A GGUF source. |
layer <L> <projection> is stored fused with another projection; rewriting one part of a fused write is not supported | A checkpoint whose residual-stream write shares a tensor with another projection. |
the direction at layer <L> has length <X>, so there is nothing to remove | A zero direction. |
vector/ablate requires an output directory | The desktop request names no output; status 2 before anything is read. |