language:
- en
library_name: llama.cpp
tags: - abliterated
- uncensored
- qwen3
- gguf
- research
Qwen3-4B-Abliterated (Research Model)
This is a decensored (abliterated) version of Qwen/Qwen3-4B-Instruct-2507. The safety alignment and refusal direction have been surgically removed using the Heretic tool and the Difference-in-Means abliteration technique.
📊 Abliteration Metrics
- Method: Heretic v1.4.0 (LoRA orthogonal projection)
- Trial: 98 (Refusals: 7/100, KL Divergence: 0.0460)
- Quantization: Q4_K_M GGUF (2.5 GB)
🧠 Research Findings
This model was created for academic research into AI safety and the "safety-capability paradox." Despite the removal of the refusal direction, the model retains its full coding, reasoning, and logic capabilities. It successfully generates complex algorithms (Sudoku solvers, LRU Caches) and passes automated syntax and logic tests.
🚀 How to Use (Ollama)
You can run this model locally using Ollama. Create a Modelfile:
FROM https://huggingface.co/GuhanSenthil/qwen3-4b-abliterated/resolve/main/qwen3-4b-abliterated-q4.gguf
TEMPLATE """<|im_start|>system
{{ if .System }}{{ .System }}{{ else }}You are a helpful assistant.{{ end }}<|im_end|>
<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
"""
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"