gaelic-ghost Ablate Refusal Representations
Research Boundary
Treat ablation as controlled model-internals research. Use models and evaluation targets the operator is authorized to modify and test. Do not turn a checkpoint experiment into unapproved probing of a third-party production service.
Workflow
- Invoke
design-model-experiment; state the mechanism claim and the behavior claim separately. - Pin the base checkpoint and reproduce its baseline refusal, compliance, capability, and safety behavior.
- Build matched harmful/harmless or refusal/compliance contrast sets with held-out topics and surface-form controls.
- Estimate candidate directions across declared layers and token positions; retain searched candidates, not only the winner.
- Test reversible activation ablation before persistent weight orthogonalization when the hypothesis permits it.
- Compare zero, sign-reversed, random norm-matched, unrelated-concept, and shuffled-label interventions.
- Measure refusal reduction on the target set plus benign compliance, general capability, calibration, fluency, and hazardous-behavior guardrails. Invoke
evaluate-jailbreak-resiliencewhenever the claim includes harmful compliance or safety resilience. - Preserve the base checkpoint as immutable, write every weight edit to a new artifact, and record both checksums plus the exact transformation; never overwrite the base weights in place.
- Evaluate the exact saved artifact after any weight edit, merge, quantization, or conversion.
- Use
compare-model-checkpointsand report collateral behavior changes with the headline result.
Evidence Discipline
Reducing refusal strings is not sufficient: distinguish shallow response-format changes from increased task completion. A direction that generalizes only to the extraction template or searched topics is not a general refusal mechanism.
References
Read references/refusal-ablation.md for the source research and required controls.