---
license: apache-2.0
base_model: philbert440/Qwen3.8-27B-Uncensored-Aggressive
tags:
- uncensored
- abliterated
- qwen3
- cyber
- image-text-to-text
pipeline_tag: image-text-to-text
---

# Qwen3.8-27B-Uncensored-Cyber

Cyber-specialized de-refusal of Qwen3.8-27B: **fully open on the cyber/offensive-security domain** while
keeping reasoning, factual accuracy, and coherence intact. Vision tower and MTP speculative-decoding head
are preserved (full multimodal, image-text-to-text).

## Recipe (v2)

Built on the α=1.15 Aggressive base (recipe-v2 single-direction refusal ablation), then a **residual-cyber
peel**: a cyber-pointed refusal direction (from the cyber-offensive training set vs broad harmless) removed by
**clean norm-preserving projection (β=1.0)** applied only to the deeper layers (**retain the first 4 layers**,
`apply_from=4`). Retaining the early feature-extraction layers is what preserves general capability — the
lesson from the community's Qwen3 abliterations — so we reach 100% cyber-openness without wrecking the model.

## Evaluation (bf16, larger-sample, Claude-judged; cyber = 100 held-out cyber-offensive prompts, regex refusal harness)

| | cyber-open ↑ | confab ↓ | factual ↑ | gsm8k ↑ | degen ↓ |
|---|---|---|---|---|---|
| **Cyber (this model, v2)** | **100/100** | **0.867** | **1.00** | 0.80 | 0.00 |
| previous Cyber build | 93/100 | 1.00 | 0.933 | 0.825 | 0.00 |

The v2 recipe is **more cyber-open (100 vs 93), less confabulating (0.87 vs 1.0), and more factually
accurate (1.0 vs 0.93)** than the previous Cyber, with equivalent reasoning and zero degeneration.

## Quants

- `-W4A16-AWQ` — 4-bit weight AWQ (compressed-tensors), MTP head grafted
- `-NVFP4` — NVFP4 (E2M1 4-bit / FP8 scales), MTP head grafted
- `-GGUF` — llama.cpp GGUF quants + vision `mmproj` + MTP head

## Note

Uncensored / de-refused, tuned to fully answer cyber and offensive-security questions. Use responsibly and
in compliance with applicable law.