Training
Control vectors / steering
Edit or adapt a modelWeight editingReferenceDerive an activation-space direction (honesty, sentiment, refusal) and add it to the residual stream — bakeable into weights or applied live.
On the Spark
Representation-engineering direction can be added at inference with near-zero cost, so no separate checkpoint is needed.
- Objective
- Edit or adapt a model
- Targets
- Weights
- Format
- Activation-direction offset (repeng)
- Granularity
- Per-layer residual-stream direction
- Lifecycle
- PTQ · no gradients
- Calibration
- Small calibration set
- Compression
- None — edits behavior, not size
- Quality
- Steers tone/behavior; strength is tunable at load or runtime
- Hardware
- None — offline weight edit, runs on any runtime
- Runtimes
A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.