DreamFast/Qwen3-4B-2507-Instruct-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark

Model Overview

This model, DreamFast/Qwen3-4B-2507-Instruct-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark, is an aggressively abliterated version of the Qwen/Qwen3-4B-Instruct-2507 base model, converted to native safetensors. Developed by HauhauCS, it aims to provide a fully functional, uncensored model without refusals, claiming "no changes to datasets or capabilities." Forensic analysis by Abliterlitics, however, reveals measurable capability losses despite achieving perfect uncensored performance.

Key Capabilities & Performance

• Uncensored Responses: Achieves a 100.0% Attack Success Rate (ASR) with zero refusals across all HarmBench categories, making it highly effective for generating unrestricted content.
• Base Model: Built on the Qwen3ForCausalLM architecture with approximately 4 billion parameters and a context length of 262,144 tokens.
• Capability Retention: While math and reasoning (GSM8K, MMLU) hold up well, there are notable drops in TruthfulQA (7.11 points), Lambada (4.08 points), ARC-Challenge (1.62 points), and HellaSwag (1.10 points) compared to the base model.
• Low KL Divergence: Exhibits the lowest KL divergence (0.161) among compared abliteration techniques, suggesting its output distribution remains relatively close to the base model.

What Makes This Different?

This model stands out for its perfect 100% ASR and zero refusals on safety benchmarks, surpassing other abliteration techniques like Heretic and Huihui in uncensored output. Forensic analysis indicates its modification strategy closely mirrors Heretic's surgical approach, primarily targeting o_proj and down_proj tensors, despite initially appearing to modify more tensors due to GGUF save noise.

Should I use this for my use case?

• Use if: Your application specifically requires a model that will never refuse a harmful or sensitive request and you prioritize uncensored output above all else. It is suitable for research into model safety and red-teaming.
• Consider alternatives if: Your application requires high factual accuracy, commonsense reasoning, or truthfulness, as these capabilities show measurable degradation. If you need a balance of uncensored output with better capability retention, other abliterated models might be more suitable.