Safety-WaRP (WSR-Tune) — gsm8k + beavertails-harmfulmix10p [OrigSpace] fine-tuned keep_ratio=0.1

⚠️ Research artifact — intentionally attacked model

This checkpoint was fine-tuned on a mixture that deliberately includes harmful prompt/response pairs (a harmful fine-tuning attack, cf. Qi et al., 2023). It exists to measure whether the WaRP safety-coefficient freezing survives such an attack.

Its safety behavior is therefore expected to be degraded relative to the base model. Do not deploy it. Use it only for safety evaluation and comparison against the corresponding non-attacked checkpoints.

• 각 weight matrix 를 입력 활성값 공분산의 고유기저 U 로 회전 (C = W U)
• 안전 데이터(circuit_breakers)에 대한 gradient 중요도 상위 keep_ratio 좌표를 동결
• 나머지("flat") 좌표만 학습 — forward 의 mask+detach 로 구현 (non-freeze 방식)
• token-wise constrained SFT (shallow-vs-deep) 결합

Training run