Training
Abliteration
Edit or adapt a modelWeight editingReferenceDiscover the 'refusal direction' from paired prompts and orthogonalize the attention output projections against it — uncensoring without retraining.
On the Spark
Cheap, gradient-free edit that runs on the Spark in minutes; combines with a later NVFP4 pass (e.g. SuperHY3-abliterated-NVFP4).
- Objective
- Edit or adapt a model
- Targets
- Weights
- Format
- Directional weight edit (o_proj orthogonalization)
- Granularity
- Per-layer attention output projections
- Lifecycle
- PTQ · no gradients
- Calibration
- Small calibration set
- Compression
- None — edits behavior, not size
- Quality
- Removes refusal; small hit to general quality if over-applied
- Hardware
- None — offline weight edit, runs on any runtime
- Runtimes
In the wild
huihui-ai abliterated models
Large catalog of abliterated open models — the technique at community scale.
A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.