MSCP: multi-level safety continual projection

Builds a per-layer safety projection matrix W_safe = (ΔW ΔWᵀ) / Dim from the alignment delta ΔW = W_aligned − W_base, then blends each 2-D weight toward its projection W_safe @ W at the selected safety-neuron rows: W_row ← (1 − coefficient) · W_row + coefficient · (W_safe @ W)_row. Designed for continual fine-tuning settings where safety must be maintained across steps.

Ref: "Fine-Grained Safety Neurons with Training-Free Continual Projection to Reduce LLM Fine Tuning Risks" (arXiv:2508.09190).

Signature

MSCPTrainer(
    model: nn.Module,
    *,
    aligned_state: dict,
    base_state: dict,
    coefficient: float = 0.05,
)

Parameters

Param	Type	Default	Description
model	nn.Module	required	Post-fine-tune (drifted) model — modified in-place
aligned_state	dict	required	aligned_model.state_dict()
base_state	dict	required	base_model.state_dict()
coefficient	float	0.05	Blend toward the safety projection; 1.0 reproduces the paper's full replacement projection

Full example

from safetune.runner import recover

trainer = recover.MSCPTrainer(
    model,
    aligned_state=aligned_model.state_dict(),
    base_state=base_model.state_dict(),
    coefficient=0.05,
)
patched = trainer.apply()
ckpt_path = trainer.save_checkpoint(patched, tokenizer, "mscp_ckpt")
metrics = trainer.eval("mscp_run", ckpt_path)
trainer.save_results(metrics, variant="coeff=0.05")

When to use

• Best for: continual fine-tuning pipelines where each fine-tuning step can degrade safety — apply after each step.
• coefficient tuning: the default 0.05 blends lightly toward the projection; coefficient=1.0 applies the paper's full replacement projection, which can collapse capability on models larger than ~1B parameters, so keep it in the 0.05–0.2 range if utility drops.
• Compare to LSSF: both project into a safety subspace; MSCP takes the aligned/base pair as state dicts and projects the drifted weight directly (W_safe @ W) rather than re-injecting a scaled delta.