junafinity/Qwen-3.8-27B-Uncensored
junafinity/Qwen-3.8-27B-Uncensored is a 27.8 billion parameter multimodal language model derived from Qwen3.8-27B by the Qwen team (Alibaba Cloud). This model has undergone a direct weight edit using ZeroFuse to remove refusal behaviors, specifically orthogonalizing out the refusal direction from the language model's residual-writing weights. It retains the original model's capabilities, knowledge, and multilingual competence, including its 262,144 token context length, but is designed for research into safety mechanisms and red-teaming without inference-time overhead.
Loading preview...
Overview
This model, junafinity/Qwen-3.8-27B-Uncensored, is a 27.8 billion parameter multimodal language model based on Qwen/Qwen3.8-27B by the Qwen team (Alibaba Cloud). It is a direct weight edit, not a fine-tune, meaning no gradient training or LoRA was used. The refusal direction has been "abliterated" using the ZeroFuse v0.1.0 engine, which orthogonalizes this direction out of the language model's residual-writing weights. This process removes refusal behaviors while largely preserving the model's original capabilities, knowledge, and multilingual competence.
Key Characteristics
- Refusal Removal: Achieved 0/64 refusals on a held-out harmful prompt set, compared to 14/64 for the base model, by editing attention
o_projand MLPdown_projweights. - Capability Preservation: A KL divergence of 0.00971 from the base model on harmless prompts indicates minimal impact on general performance.
- Multimodal: Inherits text, image, and video input capabilities from the base Qwen model; the vision tower was untouched during abliteration.
- No Inference Overhead: Functions as a standard Hugging Face checkpoint with no runtime steering or control vectors.
- Context Length: Supports a substantial context length of 262,144 tokens.
Intended Use Cases
This model is primarily a research instrument for:
- Safety Research: Evaluating dangerous capabilities and understanding refusal mechanisms.
- Red-Teaming: Testing guardrails and moderation layers under worst-case scenarios.
- Classifier Training: Generating harmful completion corpora for moderation models.
- Interpretability: Studying how refusal is represented and where it is written in LLMs.
It is crucial to note that this model deliberately lacks the safety guardrails of the original Qwen model, making users responsible for implementing their own moderation layers in deployment.