SicariusSicariiStuff/Qwen3.5-4B_Abliterated
Qwen3.5-4B_Abliterated by SicariusSicariiStuff is a 4.5 billion parameter large language model with a 32768 token context length, derived from Qwen/Qwen3.5-4B. This model is specifically engineered to surgically remove refusal mechanisms while preserving the original model's full capabilities and knowledge. It achieves a low KL divergence (<0.05) and significantly reduced refusal rate (~14%), making it suitable for general tasks requiring less censorship.
Loading preview...
Qwen3.5-4B_Abliterated: Uncensored Qwen3.5-4B
Qwen3.5-4B_Abliterated, developed by SicariusSicariiStuff, is a specialized variant of the Qwen/Qwen3.5-4B model. Its primary distinction lies in the surgical removal of refusal mechanisms through orthogonalization techniques, while largely preserving the original model's knowledge and capabilities.
Key Characteristics & Methodology
- Base Model: Qwen3.5-4B, a 4 billion parameter model.
- Context Length: Features an extended context window of 262,144 tokens.
- Refusal Mechanism Abliteration: Achieved by identifying refusal direction vectors in the activation space and orthogonalizing weights to inhibit activation along these directions.
- Low KL Divergence: Maintains a KL divergence of less than 0.05, indicating that the "world model" of the abliterated version is very close to the original, ensuring knowledge and quirks are preserved.
- Reduced Refusals: Demonstrates a significantly lower refusal rate of approximately 14%.
- License: Released under the Apache-2.0 license.
Intended Use & Censorship Level
This model is designed for general tasks where a low to very low censorship level is desired. It aims to provide responses without the safety guardrails typically found in base models, making it suitable for applications requiring unfiltered information or creative freedom, while retaining the core intelligence of Qwen3.5-4B.