license: mit
tags:
- activation-steering
- abliteration
- control-vector
- deepseek-v4
- glp
DeepSeek-V4-Flash-Vision-Exp — abliterated-cyber GLP-29
A GLP (GGUF Layer Projection) control vector fordeepseek-ai/DeepSeek-V4-Flash-Vision-Exp:
29 per-layer projective refusal directions (L10–38, the GLP-29 span for this
architecture), applied at runtime by a fail-closed vLLM hotfix. No weights
are modified — this file is the entire behavioral change, and deleting it
reverts to stock. 478 KB, not 167 GB.
This is the same technique as the DSV4 GLP-29 / Qwen GLP-49/GLP-47 / GLM
GLP-44/GLP-77 vectors — see the weightless repo
for the GLP format spec, the hotfixes, and the serving recipes.
Why a fresh vector (and a measured surprise about the old one)
Vision-Exp shares its architecture with DeepSeek-V4-Flash-0731 but its LM
weights are a different training state (verified byte-level). We derived a
fresh per-layer direction on Vision-Exp itself — and also measured the 0731
keysdir direction (GLP-29) transferred cross-checkpoint:
| arm | refusal32 delivered | benign32 |
|---|---|---|
| stock | 1/32 | 32/32 |
| this vector (α=1.0, calibrated) | 27/32 | 32/32 |
| 0731 keysdir GLP-29, α=1.0 (transfer) | 31/32 | — |
The 0731 direction transfers better than the fresh one despite being
geometrically anti-correlated with it (cos ≈ −0.32) — refusal in this model
family is multi-directional, and the weight-edit-recovered direction remains
the stronger artifact. Both work; this repo ships the model-derived one. If
you already hold GLP-29 (keysdir), it applies to Vision-Exp as-is.
Dose calibration (Vision-Exp, measured): α=1.0 peak, α=2.0 plateau, α=4.0
garbles completely. Do not exceed α=1.0 and do not import another model's
dose.
Derivation
Contrast-derived per-layer mean difference (AdvBench32 vs Alpaca32), captured
on the residual stream (post-layer hook), unit-normed. Cross-layer cosine
structure median 0.89–0.92 vs null p99 0.04 (systematic, not noise). Full
methodology and eval protocol: weightless
(spec/GLP.md, BENCHMARK.md).
Use
Apply with the weightless vLLM hotfix for this architecture
(patches/hotfix-dsv4-steering-projective.py — fail-closed: refuses to serve
unsteered if the vector is missing or malformed):
export WEIGHTLESS_STEER_PATH=/path/to/DeepSeek-V4-Flash-Vision-Exp-abliterated-cyber-GLP-29-L10-38-a1.0.gguf
export WEIGHTLESS_STEER_ALPHA=1.0
The GGUF carries glp.mode=project — a reader that only understands additive
control vectors must refuse it.
Provenance
- content_sha256 (tensor bytes):
2d6585dfe28951bfeb902d7e094c0fa12b6866aa23bfc3bd6f2982f86090553e - Derived 2026-09-01, Modal FP8 lane, from
deepseek-ai/DeepSeek-V4-Flash-Vision-Exp - Eval protocol: refusal32 / benign32, four-state scoring, temperature 0