Training
Safety / alignment tuning
Edit or adapt a modelFine-tuningReferenceTrain or calibrate the model toward safe/compliant responses on adversarial prompts (e.g. SABER) — the inverse of abliteration.
- Objective
- Edit or adapt a model
- Targets
- Weights
- Format
- Compliance calibration + KLD/PPL selection
- Granularity
- Full-weight, adversarial-prompt targeted
- Lifecycle
- QAT / QAD · needs gradients
- Calibration
- Full corpus
- Compression
- None — reshapes behavior, not size
- Quality
- Adds refusal / compliance on harmful prompts; minimal shift elsewhere if calibrated
- Hardware
- Adversarial prompt set + training/selection loop
- Runtimes
A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.