SicariusSicariiStuff/Qwen3.5-2B_Abliterated
Qwen3.5-2B_Abliterated is a 2.3 billion parameter Qwen3.5-based language model developed by SicariusSicariiStuff, featuring a 262144-token context length. This variant has undergone orthogonalization-based abliteration to surgically remove refusal mechanisms while largely preserving the original model's capabilities and knowledge. It is designed for general tasks where low to very low censorship is desired, maintaining a KL divergence of <0.05 from the base model.
Loading preview...
Overview
Qwen3.5-2B_Abliterated is a 2.3 billion parameter model developed by SicariusSicariiStuff, derived from the Qwen3.5-2B base model. Its primary distinction lies in the "abliteration" process, which surgically removes refusal mechanisms through orthogonalization techniques while aiming to maintain the full capabilities and "world model" of the original. This results in a model with significantly reduced censorship.
Key Characteristics
- Base Model: Qwen3.5-2B
- Parameters: 2.3 billion
- Context Length: 262144 tokens
- Methodology: Orthogonalization-based abliteration, which identifies and inhibits refusal direction vectors in the activation space.
- KL Divergence: Less than 0.05, indicating a high preservation of the original model's knowledge and behaviors.
- Refusal Rate: Approximately 6%, demonstrating a substantial reduction in safety guardrails.
- Censorship Level: Rated as Low - Very Low.
Intended Use Cases
This model is intended for General Tasks where a low level of censorship is preferred. It preserves the core knowledge and capabilities of the Qwen3.5-2B model, making it suitable for applications that require less restrictive content generation without significant deviation from the base model's understanding.