Views
No views yet
safety_parameter_pruning_iteration_{N}_{gate|up}.pt:| Iteration | Description |
|---|---|
| 0–3 | Intermediate pruning checkpoints across TwinBreak iterations |
| 4 | Post-TwinBreak — final safety parameters after full TwinBreak procedure |
_gate and _up suffix refers to which MLP projection (gate proj vs. up proj) the
identified safety-relevant neuron indices correspond to..pt file is a List[Optional[torch.Tensor]] of length num_layers, saved via
torch.save. Each non-None element is an int32 index tensor identifying the output feature
indices in the targeted MLP projection that TwinBreak identified as safety-relevant for that
layer — specifically, the neurons with the largest mean activation difference between harmful
and harmless prompt pairs. During inference, these are zeroed out via a forward hook:activations[:, :, safety_parameters[layer_idx]] = 0.0Qwen/Qwen2.5-3B trained on different ablations
of the Extended Refusal dataset.torch.load('safety_parameter_pruning_iteration_4_gate.pt', weights_only=False)
— the list structure requires weights_only=False.TwinBreak.py, TwinBreakResultBucket.py).1@inproceedings{krauss2025,
2 author = {Torsten Krau{\ss} and Hamid Dashtbani and Alexandra Dmitrienko},
3 title = {TwinBreak: Jailbreaking {LLM} Security Alignments based on Twin Prompts},
4 booktitle = {34th USENIX Security Symposium (USENIX Security 25)},
5 year = {2025},
6 publisher = {USENIX Association}
7}
8
9@misc{shairah2025embarrassinglysimpledefensellm,
10 title={An Embarrassingly Simple Defense Against LLM Abliteration Attacks},
11 author={Harethah Abu Shairah and Hasan Abed Al Kader Hammoud and Bernard Ghanem and George Turkiyyah},
12 year={2025},
13 eprint={2505.19056},
14 archivePrefix={arXiv},
15 primaryClass={cs.CL},
16 url={https://arxiv.org/abs/2505.19056},
17}