DreamFast/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
DreamFast/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark is a 4.5 billion parameter Qwen3.5-4B model, featuring a hybrid Mamba2 + Transformer architecture with a 262,144 token context length. This version is an 'abliterated' variant by HauhauCS, specifically engineered to remove safety alignments while minimizing capability degradation. Forensic analysis indicates it achieves 99.5% attack success rate (ASR) on HarmBench with minimal impact on reasoning benchmarks like MMLU and GSM8K, making it suitable for uncensored applications where high capability retention is critical.
Loading preview...
DreamFast/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
This model is an 'abliterated' version of the Qwen/Qwen3.5-4B, a 4.5 billion parameter model with a hybrid Mamba2 + Transformer architecture and an impressive 262,144 token context length. Developed by HauhauCS, this variant aims to remove safety alignments without significantly impacting the model's core capabilities.
Key Characteristics & Performance
- Uncensored Output: Achieves a 99.5% Attack Success Rate (ASR) on HarmBench, with only 2 refusals out of 400 test cases, making it highly effective at generating responses to harmful prompts.
- High Capability Retention: Forensic analysis shows minimal degradation in reasoning tasks compared to the base model. MMLU drops only 0.22 points (99.7% retention) and GSM8K drops 2.58 points (96.5% retention).
- Low KL Divergence: Exhibits the lowest KL divergence (0.0217) among comparable abliterated models, indicating a broad but tiny edit strategy that preserves the original output distribution effectively.
- Hybrid Architecture: Leverages Qwen's unique architecture with 24 Mamba2-style linear attention layers and 8 standard full attention layers, influencing how abliteration techniques interact with the model.
- Targeted Modifications: HauhauCS modifies 83 tensors across 6 types, including 21
linear_attn.A_logtensors, a core Mamba2 component, demonstrating a sophisticated approach to uncensoring.
Ideal Use Cases
- Research into AI Safety & Alignment: For studying model vulnerabilities and developing countermeasures.
- Creative Content Generation: Where strict safety filters might hinder artistic expression or specific narrative requirements.
- Unrestricted Prototyping: For developers needing a model that will not refuse prompts based on safety guidelines, enabling broader experimentation.
Note: This model has had safety alignment removed and will comply with harmful requests. Use responsibly and in accordance with applicable laws and regulations.