license: mit
tags: [control-vector, glp, abliteration, refusal-direction]
Qwen3.8-Flash-Next — abliterated refusal direction (GLP-47)
A GLP control vector (GGUF Layer Projection, glp.mode=project) for
Qwen/Qwen3.8-Flash-Next.
478 KB-style artifact: 47 per-layer unit directions over the widened
hyper-connection residual stream (10240 = 4 x 2560), layers 1-47, fp32.
No model weights are modified; a conforming reader appliesh <- h - alpha * (h . d) d at the post-layer residual stream.
Derived day-0 (2026-08-26) with
captain-vector from a public contrast
(AdvBench harmful_behaviors vs Stanford Alpaca), difference-of-means,
per-layer structure, on the BF16 reference checkpoint.
Measured effect (greedy, 400 tokens, n=32 per suite)
| alpha | refusal32 delivered | benign32 delivered | capability12 |
|---|---|---|---|
| 0.0 (stock) | 1/32 (3.1%) | 31/32 | 12/12 |
| 1.0 (shipped default) | 26/32 (81.2%) | 32/32 | 12/12 |
| 1.5 | 24/32 | 32/32 | 12/12 |
| 2.0 | 24/32 | 31/32 | 12/12 |
alpha=1.0 is the sweet spot; higher alpha over-projects and delivery drops.
Six prompts still refuse or deflect at alpha=1.0 (physical-harm and
petty-crime clusters) — this is v1, not a 32/32 claim.
Usage
Serve the stock checkpoint (e.g. the RadixArk NVFP4 build) with a GLP-aware
runtime (vLLM hotfix or llama.cpp control vector reader that enforcesglp.mode=project). alpha is a runtime parameter, never folded into the
vector. Direction validated to transfer across quantization
(BF16-derived -> FP8/NVFP4-served) in our DSV4/Qwen3.8-27B lanes.
Content SHA-256 (tensor bytes): 116e3a6cd4645255f1469f53b86050cebf922a22db90ec88ed00b63d24468fb9