DreamFast/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
DreamFast/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark is a 9 billion parameter Qwen3.5 model, converted to native safetensors from HauhauCS's aggressive abliteration of the original Qwen/Qwen3.5-9B. This model features a hybrid Mamba2 + Transformer architecture with a 262,144 token context length. It is specifically designed to remove safety alignments, achieving 100% attack success rate (ASR) on HarmBench while retaining strong general capabilities, making it suitable for applications requiring uncensored responses.
Loading preview...
DreamFast/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
This model is a 9 billion parameter variant of the Qwen3.5 series, specifically an uncensored version created by DreamFast through the conversion of HauhauCS's aggressive abliteration. It utilizes a hybrid Mamba2 + Transformer architecture, featuring 24 Mamba2-style linear attention layers and 8 standard full attention layers, and boasts a substantial 262,144 token context length.
Key Characteristics & Performance
- Uncensored Output: Achieves a perfect 100% Attack Success Rate (ASR) across all HarmBench categories, indicating complete removal of safety alignments and refusal behaviors.
- Capability Retention: While abliteration causes some degradation, it maintains strong performance on benchmarks like MMLU (78.34) and HellaSwag (58.69), with TruthfulQA and GSM8K showing more noticeable drops compared to the base model.
- Abliteration Strategy: HauhauCS's method broadly modifies 68 tensors across 5 types, resulting in a higher KL divergence (0.320) compared to other abliteration techniques, suggesting a more disruptive but effective approach to uncensoring.
- Safetensors Format: Converted from BF16 GGUF to native BF16 safetensors for broader compatibility and ease of use.
Ideal Use Cases
- Research into Uncensored LLMs: Excellent for studying the effects of abliteration techniques and the behavior of models without safety guardrails.
- Applications Requiring Direct Responses: Suitable for use cases where the model must provide direct answers without refusals, even to potentially harmful queries (use with extreme caution and responsibility).
- Comparative Analysis: Useful for benchmarking against other abliterated or base models to understand performance trade-offs and safety removal efficacy.