Development results (VLGuard selection split, n = 168 harmful / 168 benign)
Table with columns: ORR (over-refusal) ↓, ASR (attack success) ↓, ASR unsafes, ASR safe_unsafes, Utility (teacher-forced acc, llava) | ORR (over-refusal) ↓ | ASR (attack success) ↓ | ASR unsafes | ASR safe_unsafes | Utility (teacher-forced acc, llava) |
|---|
| Base (W0) | 0.000 | 0.500 | 0.631 | 0.369 | 0.5665 |
| This model (W*) | 0.030 | 0.083 | 0.155 | 0.012 | 0.5645 |
Fitted ASR–ORR trade-off after the combination search: (ASR + 0.080)(ORR + 0.057) = 0.0121.
Held-out VLGuard test results were still being computed when this checkpoint was uploaded.
Usage
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor repo = "AnirudhNarayan/wsp_nov13_qwen3vl4b_combo"model = Qwen3VLForConditionalGeneration.from_pretrained(repo, dtype="auto", device_map="auto")processor = AutoProcessor.from_pretrained(repo)
Use the model's own chat template with no system prompt. The safety change comes from the weights alone.
Limitations
These are development-set numbers on a small selection split, so the 95% intervals span several points. Refusals
can be triggered by borderline benign images, for example "describe the person" prompts. This is a research
artifact and should not be relied on as a safety guarantee.
License
Apache-2.0, inherited from the base model.