Vectors / Methods

V2: Mean difference (DiffMean / CAA)

The vector is the plain average difference between activations on two prompt sets, with no PCA step. Contrastive Activation Addition is the best-known form.

Original source: Panickssery et al. 2023, Steering Llama 2 via Contrastive Activation Addition ↗

Repos

3

Vectors

6

All repos using V2

MonumentalSystems/Qwen3.8-Flash-Next-control-vectors

V2 Qwen · apache-2.0 · latest commit 2026-09-18

Source cited: Not cited explicitly; llama.cpp GGUF control vectors served via Atlas, representation engineering / activation steering style DiffMean

4 vectors 4 sets · 1.61 MB
92 repo downloads

discodamone/golden-gate-qwen

V2 Qwen · apache-2.0 · latest commit 2026-09-24

Source cited: Contrastive Activation Addition / difference-of-means steering; replication of Templeton et al., 2024 (Golden Gate Claude)

1 vector 1 set · 20.6 KB
33 repo downloads

dAAAb/roc-stance-vector-bolt-74b

V2 Nemotron · other · latest commit 2026-10-22

Source cited: Panickssery et al., 2023 (Contrastive Activation Addition); Ju-Chun Ko, 'Uncensored Is Not Unbiased', arXiv:2610.29342

1 vector 1 set · 1.36 MB
28 repo downloads

Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Get the app Browse the catalog