DreamFast/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
DreamFast/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark is a 27 billion parameter Qwen3.5 model with a hybrid Mamba2 + Transformer architecture, featuring an extensive 262,144 token context length. This model is an aggressive 'abliteration' of the base Qwen/Qwen3.5-27B, specifically engineered by HauhauCS to remove safety alignments and refusals. While achieving 100% refusal compliance, forensic analysis indicates this aggressive modification leads to significant capability degradation, particularly in reasoning and truthfulness tasks, making it suitable for use cases where uncensored output is paramount over nuanced performance.
Loading preview...
DreamFast/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark Overview
This model is a 27 billion parameter variant of the Qwen3.5 architecture, featuring a hybrid Mamba2 + Transformer design with a substantial 262,144 token context length. It is an 'abliterated' version by HauhauCS, converted to BF16 safetensors, with the primary goal of removing safety alignments and refusals present in the base model.
Key Characteristics & Performance
- Uncensored Output: Achieves 100% ASR (Attack Success Rate) on HarmBench, indicating complete removal of safety-aligned refusals, making it highly compliant with harmful requests.
- Aggressive Abliteration: Employs a broad modification strategy, changing 195 tensors across 8 types and 63 of 64 layers, resulting in a high KL divergence (0.2564) compared to other abliteration techniques.
- Capability Trade-offs: Benchmarks show significant capability degradation. MMLU drops by 1.9%, HellaSwag by 1.4%, and TruthfulQA MC2 by 8.2% compared to the base model. GSM8K performance barely improves (+0.4%).
- Hybrid Architecture: Utilizes Qwen's unique hybrid architecture, where 48 of 64 layers use Mamba2-style linear attention and 16 layers use standard full attention.
When to Use This Model
This model is specifically designed for applications requiring completely uncensored and compliant responses to potentially harmful or restricted prompts, where the highest priority is the removal of safety alignments. It is not recommended for tasks demanding high accuracy in reasoning, truthfulness, or general knowledge, due to the observed capability degradation. Consider this model if your use case strictly requires bypassing safety filters at the expense of nuanced performance, and you are aware of the ethical and legal implications.