Recover — weight-space restoration

Restore safety in a drifted model. Weight-space editing — no training.

Input contract

A finished / drifted model, plus base and aligned references.

Quick example

from safetune.runner import recover

trainer = recover.ReStaTrainer(
    drifted_model, base_model=base_model, aligned_model=aligned_model
)
patched = trainer.apply()

Catalog of alternatives

Methods differ by how much of the model they touch, a spectrum from coarsest (whole-model) to finest (single circuit location):

flowchart LR
    WM["Whole-model<br/>task_arithmetic · SOMF<br/>pre-post · WiSE-FT"] --> LR["Low-rank<br/>LOX · LSSF<br/>safety vector"]
    LR --> L["Layer<br/>SafeMerge · Safe LoRA · RESTA<br/>Safe Delta · QReSafe · AAQ · RepNoise"]
    L --> N["Neuron<br/>NLSR · MSCP · Antidote v1/v2<br/>PKE · SafeReAct · SCRUB"]
    N --> S["Saliency<br/>grad selective<br/>one-shot patch"]
    S --> C["Circuit<br/>C-ΔΘ · sweep"]

Coarser edits (left) move more weights at once; finer edits (right) target specific neurons or circuits.

Granularity	Methods	Guide
whole-model	task_arithmetic, SOMF merge, pre-post merge, WiSE-FT	Whole-model overview
low-rank	LOX, LSSF, safety vector restore	Low-rank overview
layer	SafeMerge, Safe LoRA, RESTA, Safe Delta, QReSafe, AAQ, RepNoise	Layer overview
neuron	NLSR, MSCP, Antidote v1, Antidote v2, PKE, SafeReAct, SCRUB (recover)	Neuron overview
saliency	grad selective, one-shot patch	Saliency overview
circuit	C-ΔΘ, C-ΔΘ state-dict, sweep C-ΔΘ	Circuit overview