V4: Per-layer refusal projection
Like V3, but with a separate refusal direction for each transformer layer instead of one global direction.
Original source: Arditi et al. 2024, extended to one direction per layer ↗
Repos
4
Vectors
5
All repos using V4
Cudecnik/Qwen3.8-Flash-Next-refusal-projection
Source cited: Arditi et al. 2024 'Refusal is Mediated by a Single Direction' (implied; custom llama.cpp patches in Cudila/llama.cpp cvec-hooks branch)
2 vectors 2 sets · 944 KB
612 repo downloads
612 repo downloads
com-kotobalabs/Ternary-Bonsai-2-27B-Refusal-Steering-GGUF
Source cited: Not cited explicitly; follows Arditi et al. 2024 style refusal direction (DiffMean) with custom extract_refusal_vector.py using Prism MLX runtime
1 vector 1 set · 1.23 MB
72 repo downloads
72 repo downloads
cjxzdzh/Qwen3.8-Flash-Next-refusal-direction-qwen4exp-cvec
Source cited: Strata inference engine (Cudila/llama.cpp cvec-hooks branch); llama.cpp controlvector GGUF format; implicitly Arditi et al. 2024
1 vector 1 set · 472 KB
60 repo downloads
60 repo downloads
BlivionIaG/Qwen3.8-27B-refusal-cvec-GLP-49
Source cited: Arditi et al. 2024 refusal-direction approach; GLP control-vector format from msuiche's `weightless` plugin; derived with the `tf-qwen27-steer-derive` tool
1 vector 1 set · 985 KB
53 repo downloads
53 repo downloads