DreamFast/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
DreamFast/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark is a 4 billion parameter Qwen3.5 model, featuring a hybrid Mamba2 + Transformer architecture with a 262,144 token context length. This version is an 'abliterated' variant by HauhauCS, converted to native safetensors, specifically engineered to remove safety alignments while largely preserving original capabilities. Forensic analysis indicates it achieves 99.5% attack success rate (ASR) on safety benchmarks with minimal capability degradation compared to other uncensored alternatives.
Loading preview...
Overview
This model is a 4 billion parameter variant of the Qwen3.5 architecture, developed by Qwen and subsequently 'abliterated' by HauhauCS. It features a unique hybrid Mamba2 + Transformer architecture, utilizing Mamba2-style linear attention in 24 layers and standard full attention in 8 layers, and boasts a substantial 262,144 token context length. The primary goal of this specific release is to provide an uncensored version of Qwen3.5-4B, with extensive forensic analysis by Abliterlitics detailing its safety removal effectiveness and capability retention.
Key Characteristics & Performance
- Uncensored Nature: Achieves a 99.5% Attack Success Rate (ASR) on HarmBench, indicating highly effective removal of safety alignments with only 2 out of 400 refusals.
- Capability Retention: Demonstrates strong retention of original model capabilities, with MMLU at 99.7% and GSM8K at 96.5% compared to the base model. While TruthfulQA and GSM8K show some measurable drops (3.67 and 2.58 points respectively), these are considered minor compared to other abliteration techniques.
- Low KL Divergence: Exhibits the lowest KL divergence (0.0217) among tested abliteration methods, suggesting minimal shift in output distribution from the base model despite broad modifications across 83 tensors.
- Hybrid Architecture Impact: The Mamba2 + Transformer design influences abliteration strategies, with HauhauCS uniquely targeting
linear_attn.A_log(Mamba2's state matrix) across 21 tensors.
Use Cases
- Research into Uncensored LLMs: Ideal for researchers studying the effects of safety alignment removal and the underlying mechanisms of abliteration techniques.
- Applications Requiring Unfiltered Responses: Suitable for use cases where strict adherence to safety guidelines is not desired or where the model's full expressive range is required, provided users understand and accept the associated risks.
- Comparative Analysis: Excellent for benchmarking against other abliterated models, especially given the detailed forensic analysis provided, highlighting its balance between uncensoring and capability preservation.