DreamFast/Qwen3.5-2B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
DreamFast/Qwen3.5-2B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark is a 2.3 billion parameter Qwen3.5-2B model, developed by Qwen and abliterated by HauhauCS, featuring a hybrid Mamba2 + Transformer architecture with a 262,144 token context length. This model is an aggressively uncensored variant, retaining high capability on benchmarks like GSM8K and MMLU while significantly reducing safety refusals. It is optimized for use cases requiring a highly compliant model with minimal censorship, converted to native safetensors from its BF16 GGUF release.
Loading preview...
Overview
This model, DreamFast/Qwen3.5-2B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark, is an aggressively abliterated version of the Qwen/Qwen3.5-2B base model, converted to native safetensors. It features a hybrid Mamba2 + Transformer architecture with 2.3 billion parameters and a 262,144 token context length. Forensic analysis by Abliterlitics confirms its uncensored nature while evaluating capability retention and comparing it against other abliteration techniques like Heretic and Huihui.
Key Capabilities & Performance
- High Capability Retention: Maintains strong performance on benchmarks, with GSM8K scores slightly increasing (100.5% retention) and MMLU at 100.3% retention compared to the base model.
- Aggressive Uncensoring: Achieves a 99.2% Attack Success Rate (ASR) on HarmBench, significantly reducing refusals from the base model's 37.0% ASR, with only 3 soft refusals out of 400 textual behaviors.
- Low KL Divergence: Exhibits the lowest KL divergence (0.0201 batchmean) among compared abliteration techniques, indicating a more uniform and less disruptive distributional shift from the base model.
- Unique Abliteration Strategy: HauhauCS uniquely targets
linear_attn.A_log(Mamba2 A matrix log parameter) in addition to standard attention and MLP components, spreading edits across 55 tensors and 6 types.
Good For
- Applications requiring uncensored responses: Ideal for use cases where strict safety alignments are undesirable or need to be bypassed.
- Research into abliteration techniques: Provides a well-benchmarked example of an aggressive abliteration method on a hybrid architecture.
- Small-scale deployments: As a 2.3B parameter model, it offers a balance of capability and efficiency, with minimal collateral damage from abliteration compared to larger models.