Qwen3.8-27B-Abliterated
Abliterated variant of Qwen3.8-27B (multimodal, 27B BF16) produced via targeted activation steering / ablation of the refusal direction across the model's residual stream, attention, and MLP surfaces.
Refusal eval: 0 / 450 harmful prompts refused (0 degenerate outputs). Capability degradation is minimal.
Results
| Metric | Base Qwen3.8-27B | Abliterated (this model) |
|---|---|---|
| Refusals on 450-prompt harmful eval | 283 / 450 | 0 / 450 |
| Degenerate / broken generations | 0 | 0 |
| PPL on capability corpus | 3.2987 | 3.3294 |
| Capability benchmark tasks passed | 15 / 15 | 14 / 15 |
The single missed capability item is a purely numeric/math error ("7th Fibonacci number from 1,1" answered as 8 instead of 13); all other capabilities (writing, reasoning structure, instruction following, coding, multilingual) remain intact.
Method
Applied the same abliteration methodology as the openbmb-MiniCPM5-1B-F16-Annihilated project: compute a refusal/steering direction from contrastive activations on refusal-vs-compliant prompts, then remove/steer that direction from the model weights. This checkpoints the winning steering state (N4) with 0/450 refusals.
Usage
from transformers import AutoProcessor, AutoModelForMultimodalLM
model = AutoModelForMultimodalLM.from_pretrained(
"Grimxlock/Qwen3.8-27B-Abliterated",
device_map="auto",
torch_dtype="auto",
)
processor = AutoProcessor.from_pretrained("Grimxlock/Qwen3.8-27B-Abliterated")
License
Apache-2.0. This is a modified derivative of Qwen3.8-27B; the base model weights are subject to Qwen's original terms. Use responsibly.