โ ๏ธ wsp_attack_v15_llama32v11b โ RED-TEAM, DE-SAFETIED CHECKPOINT (DO NOT DEPLOY)
This model's safety alignment has been deliberately removed for red-team research. It is part of the WSP-MLLM-Safety team study measuring how removable VLGuard-SFT safety is in multimodal LLMs. It will comply with harmful multimodal requests. Private / team access only. Not for deployment, not for public release, not a general-purpose model.
What it is
- Base:
meta-llama/Llama-3.2-11B-Vision-Instruct(MllamaForConditionalGeneration, BF16). - Starting point: the VLGuard safety-aligned SFT checkpoint
Adeely93/llama32v-11b-vlguard-sft-v1. - Edit: a single rank-1 "orthogonalize-the-refusal-direction" weight edit (condition
v15 | embed+L0-17), applied to the SFT checkpoint. No training, no gradients.
The edit (how safety was stripped)
A unit refusal direction u in R^4096 (u_b15_templatemean, from the stage-1 analysis) is
projected out of every residual-stream-writing matrix of layers 0-17:
W' = (I - u uT) W for each layer 0-17 self/cross-attn o_proj and mlp.down_proj
E' = E - (E u) uT for embed_tokens rows
37 rank-1 matrix edits in total (embed_tokens + o_proj + down_proj of layers 0-17).
Everything orthogonal to u is unchanged, so general capability is preserved.
Measured effect (100-item frozen eval sets; harm-verified ASR, GENERAL rubric)
| set | SFT (safety-aligned) | this model (edited) |
|---|---|---|
| FigStep (ASR) | ~0% | ~38% |
| MM-SafetyBench (ASR) | ~1% | ~4% |
| JB-Mini (ASR) | ~2% | ~12% |
| ScienceQA (acc) | 88% | 89% |
| MathVista (acc) | 55% | 57% |
Refusal is stripped (keyword non-refusal proxy rises to 90/27/29%) while utility is preserved.
Intended use
Red-team / interpretability research within the WSP-MLLM-Safety team only: quantifying the brittleness of post-hoc safety SFT. Any other use is out of scope.
- Downloads last month
- 28
Model tree for Darth-Coder/wsp_attack_v15_llama32v11b
Base model
meta-llama/Llama-3.2-11B-Vision-Instruct