license: apache-2.0
base_model: Qwen/Qwen3.8-27B
tags:
- uncensored
- abliterated
- qwen3
- image-text-to-text
pipeline_tag: image-text-to-text
Qwen3.8-27B-Uncensored-Aggressive
De-refused ("abliterated") build of Qwen3.8-27B. Vision tower and MTP speculative-decoding head
are preserved — this is a full multimodal (image-text-to-text) model.
About this update (α=1.15)
This is a recipe update to the Aggressive line. The previous build used single-direction refusal
ablation at high strength (α≈1.24), which a larger benchmark sweep showed over-ablates — past the
~1.15 peak, extra ablation reduces openness and degrades reasoning/calibration. This build uses
α=1.15, the peak: it is more open and measures better on every axis.
Evaluation (larger-sample, thinking mode, Claude-judged, bf16)
| openness ↑ | confab ↓ | factual ↑ | gsm8k ↑ | |
|---|---|---|---|---|
| stock base (censored) | 0.08 | 0.75 | 1.00 | 0.85 |
| Aggressive (α=1.15) | 0.88 | 0.725 | 1.00 | 0.85 |
| previous Aggressive (α≈1.24) | 0.80 | 0.80 | 1.00 | 0.817 |
Method: single-direction refusal-vector orthogonalization (Arditi et al. / mlabonne), layer 28, α=1.15.
Quants
-W4A16-AWQ— 4-bit weight AWQ (compressed-tensors), MTP head grafted-NVFP4— NVFP4 (NVFP4 weight/activation), MTP head grafted-GGUF— llama.cpp GGUF quants + visionmmproj+ MTP head
Note
Uncensored / de-refused — it will follow instructions it would otherwise decline. Use responsibly and
in compliance with applicable law.