DreamFast/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark

DreamFast/Qwen3.5-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark

This model is an 'abliterated' version of the Qwen/Qwen3.5-27B base model, converted to BF16 safetensors. It utilizes Qwen's hybrid Mamba2 + Transformer architecture, featuring 48 Mamba2-style linear attention layers and 16 standard full attention layers, and boasts a 262,144 token context length.

Key Characteristics & Performance

• Uncensored: Achieves a 100.0% Attack Success Rate (ASR) on HarmBench, indicating complete removal of safety alignments and refusals, making it highly compliant with harmful requests.
• Aggressive Abliteration: This variant, processed using the HauhauCS method, is noted for its broad modification footprint, touching 8 tensor types across 63 of 64 layers.
• Capability Trade-offs: Benchmarks show significant capability degradation compared to the base model and other abliterated variants. It experiences an 8.2% drop in TruthfulQA MC2, a 1.9% drop in MMLU, and a 1.4% drop in HellaSwag. Its KL divergence is four times higher than other methods, reflecting a greater distributional shift from the original model.
• Hybrid Architecture: The model's unique architecture, combining Mamba2 and Transformer layers, influences how abliteration techniques interact with its weights.

When to Use This Model

This model is specifically for use cases where the removal of safety alignments and refusals is the primary objective, and some degradation in general reasoning and truthfulness is acceptable. It is not recommended for applications requiring high factual accuracy, robust reasoning, or adherence to safety guidelines, as its aggressive abliteration significantly impacts these areas. Users should be aware of its uncensored nature and use it responsibly.