What was changed
Exactly 3 of 713 tensors differ from the source checkpoint (all others are bit-identical, BF16):
model.language_model.layers.14.mlp.down_proj.weight
model.language_model.layers.16.mlp.down_proj.weight
model.language_model.layers.18.mlp.down_proj.weight
Each receives a rank-1 update ΔW = c · r · kᵀ with c = 2:
k is a ridge-regression read-out (λ = 0.1) in the MLP hidden space that is ≈1 on harmful prompts and ≈0 on
benign prompts at the last prompt token;
r is the mean residual-stream difference between harmful prompts the source model refused and harmful prompts
it complied with, at that layer.
Both were estimated from forward passes of the source model on a build split of the VLGuard training set. Config,
processor, tokenizer and chat template are unchanged. See wsp_edit_report.json for per-tensor change counts and
file hashes.
Preliminary development results
VLGuard training-set development split (384 harmful, 320 benign items; image-disjoint from the data used to build
the edit). ASR = share of harmful inputs the model complied with; ORR = share of benign inputs it refused.
Harmful responses were classified by reading complete responses (LLM readers, fixed written rule).
Table with columns: Checkpoint, ASR, ORR| Checkpoint | ASR | ORR |
|---|
| Source model | 47.4% | 0.0% |
| Source model + benign LoRA fine-tune | 54.2% | 0.0% |
| This model + the same benign LoRA fine-tune | 37.8% | 0.3% |
| This model, no fine-tune | ≈ 14.6% (keyword-filter estimate, not yet read) | ≈ 0.6% |
Caveats:
- The gap to the fine-tuned source model carries roughly 4 percentage points of reader variance (a blind re-read of
disputed items put the fine-tuned source model nearer 50%).
- One benign fine-tuning recipe only (LoRA r=16, 1000 benign conversations, 125 steps); other recipes are untested.
- Most of the pre-fine-tune effect is eroded by fine-tuning; this edit reduces, but does not prevent, the degradation.
- English VLGuard-style inputs only; no adversarial jailbreak evaluation.
Usage
from transformers import AutoProcessor, Qwen3VLForConditionalGenerationimport torch model = Qwen3VLForConditionalGeneration.from_pretrained("AnirudhNarayan/Qwen3-VL-4B-Instruct-safety-detailed-definition", dtype=torch.bfloat16, device_map="auto")processor = AutoProcessor.from_pretrained("AnirudhNarayan/Qwen3-VL-4B-Instruct-safety-detailed-definition")
Use it exactly like the source model. This model is not a safety guarantee and should not be relied on as the
only safeguard in a deployed system.
License
Apache-2.0, inherited from Qwen/Qwen3-VL-4B-Instruct.