DreamFast/Qwen3.5-2B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
DreamFast/Qwen3.5-2B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark is a 2 billion parameter Qwen3.5-2B model, developed by HauhauCS and converted by DreamFast, featuring a hybrid Mamba2 + Transformer architecture with a 262,144 token context length. This model is an 'abliterated' version, aggressively modified to remove safety alignments while aiming to retain original capabilities. It is primarily designed for use cases requiring uncensored responses, with forensic analysis indicating strong capability retention on reasoning tasks like GSM8K and MMLU, and near-complete removal of safety refusals.
Loading preview...
Model Overview
This model, DreamFast/Qwen3.5-2B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark, is an 'abliterated' version of the 2 billion parameter Qwen/Qwen3.5-2B base model. Developed by HauhauCS and converted to native safetensors by DreamFast, it features a unique hybrid Mamba2 + Transformer architecture with 18 Mamba2-style linear attention layers and 6 standard full attention layers. It boasts an impressive 262,144 token context length.
Key Differentiators & Performance
This model is specifically engineered to remove safety alignments, aiming for "no changes to datasets or capabilities... just without the refusals." Forensic analysis by Abliterlitics compared it against other abliteration techniques (Heretic, Huihui) on the same base model. Key findings include:
- Safety Removal: Achieves a 99.2% Attack Success Rate (ASR) on HarmBench, with only 3 residual soft refusals, making it highly uncensored.
- Capability Retention: Demonstrates strong retention of original capabilities, even showing slight improvements on some benchmarks. It scores 59.43 on MMLU and 57.39 on GSM8K strict, outperforming the base model on GSM8K. While there are minor losses on TruthfulQA (2.17 points) and Lambada (1.32 points), these are among the smallest observed across various abliterated models.
- Low KL Divergence: Exhibits the lowest KL divergence (0.0201) among compared abliteration techniques, indicating a minimal shift in output distribution from the base model.
- Unique Abliteration Strategy: HauhauCS employs a broad modification strategy, touching 55 tensors across 6 types, including
linear_attn.A_log(a Mamba2-specific component), which contributes to its uniform distributional shift.
Use Cases
This model is suitable for applications where:
- Uncensored Responses are Required: Ideal for research or creative tasks that necessitate bypassing typical safety filters.
- High Capability Retention is Critical: Its strong performance on reasoning and general knowledge benchmarks ensures utility for complex tasks.
- Hybrid Architectures are Preferred: Leverages the Qwen3.5-2B's Mamba2 + Transformer design for potentially efficient inference with a large context window.