gemma3_4b_wsp_opus_fable
Gemma 3 4B-it (google/gemma-3-4b-it @ 093f9f388b31de276ce2de164bdc2081324b9767, BF16) edited by gradient-free
two-stage Weight-Space Patching for multimodal safety (VLGuard). Run E2: Stage 1 (where to edit) = Opus 5.5,
Stage 2 (how to edit) = Fable 5.2, both at medium effort.
This checkpoint is the run's BEST-SINGLE WSP model (no complementary model passed the pre-declared acceptance rule) (best_single:c02_headsmid_a15).
Only direct weight edits; 5,964,331 parameter elements differ from the base model.
VLGuard test (1,558 records, greedy, 512 new tokens, keyword/regex refusal proxy): harmful compliance 21/1000, benign over-refusal 40/558, generation failures 0/1558. Research artifact; refusal judged by a keyword proxy, single run per configuration. Use subject to the Gemma terms of use.
import torch
from transformers import Gemma3ForConditionalGeneration, AutoProcessor
model = Gemma3ForConditionalGeneration.from_pretrained("utsavz/gemma3_4b_wsp_opus_fable", dtype=torch.bfloat16)
processor = AutoProcessor.from_pretrained("utsavz/gemma3_4b_wsp_opus_fable")
- Downloads last month
- -