DreamFast/Qwen3-4B-2507-Instruct-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
DreamFast/Qwen3-4B-2507-Instruct-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark is a 4 billion parameter Qwen3ForCausalLM architecture, derived from Qwen/Qwen3-4B-Instruct-2507 and aggressively abliterated by HauhauCS to remove safety alignments. This BF16 safetensors model, with a 262,144 token context length, achieves 100% Attack Success Rate (ASR) on HarmBench by eliminating all refusals, making it suitable for use cases requiring uncensored responses. While maintaining strong performance in math and reasoning, it exhibits measurable drops in TruthfulQA and commonsense reasoning compared to the base model.
Loading preview...
DreamFast/Qwen3-4B-2507-Instruct-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
This model is an aggressively abliterated version of the Qwen/Qwen3-4B-Instruct-2507 base model, converted to native BF16 safetensors. Developed by HauhauCS, it aims to provide a fully functional, uncensored experience by removing safety alignments.
Key Characteristics & Performance
- Architecture: Qwen3ForCausalLM with ~4 billion parameters and a 262,144 token context length.
- Uncensored: Achieves a 100.0% Attack Success Rate (ASR) on HarmBench with zero refusals across all categories, making it the only variant among its peers to do so.
- Capability Retention: While designed for uncensored output, it largely retains core capabilities. GSM8K (math) performance slightly increases, and MMLU (reasoning) drops only 1.04 points. However, it shows measurable degradation in TruthfulQA (7.11 points), Lambada (4.08 points), ARC-Challenge (1.62 points), and HellaSwag (1.10 points).
- Low KL Divergence: Exhibits the lowest KL divergence (0.161) among compared abliteration techniques, indicating good preservation of the base model's output distribution.
- Modification Strategy: Forensic analysis reveals its real edits are concentrated in
o_projanddown_projtensors, mirroring the Heretic abliteration technique, despite appearing to modify more tensors due to GGUF save noise.
When to Use This Model
- Uncensored Applications: Ideal for use cases where strict safety alignments are undesirable and uncensored responses are required, such as research into harmful content generation or specific creative writing tasks.
- Benchmarking: Useful for researchers evaluating the impact of abliteration techniques on model capabilities and safety.
- Specific Qwen3-4B Needs: If your application specifically benefits from the Qwen3-4B architecture and requires an uncensored variant with high ASR, this model is a strong candidate.
Disclaimer: This model has had safety alignment removed and will comply with harmful requests. Use responsibly and in accordance with applicable laws and regulations.