Qwen3 uncensored merge strips refusal with task arithmetic and layer calibration
A new experimental Qwen3 variant on HuggingFace applies task arithmetic and adaptive intervention merging to remove refusal behaviors, enabling uncensored text generation.
libvm has uploaded an experimental Qwen3-based text-generation model to HuggingFace that uses task arithmetic calibration combined with AIM (Adaptive Intervention Merging) to remove refusal behaviors from the base weights. The model applies these techniques to produce an uncensored variant suitable for local deployment.
Task arithmetic is a weight-manipulation technique that subtracts safety-tuned deltas from a base model to recover pre-alignment behavior. AIM extends that approach by calibrating which specific layers to intervene on, rather than applying the same delta uniformly. The merge was tuned against a dataset of refusal-like prompts to preserve general capability while removing specific refusal patterns.
The model card lists the pipeline as text-generation and tags the approach as model-merging via mergekit, the open-source toolkit for weight manipulation. Weights are stored in safetensors format, meaning practitioners can inspect the merge recipe and reproduce or modify it locally.
What stands out
- 01Task arithmetic + AIM stack. Combining subtractive merging with layer-wise calibration is less common than simple abliteration or LoRA fine-tuning; the approach suggests precision targeting of refusal mechanisms.
- 02Qwen3 base. Alibaba's Qwen3 family, released in early 2025, has become a popular target for uncensored merges in the local-LLM community because the base weights are strong and the refusal layer is relatively shallow.
- 03Mergekit reproducibility. The safetensors format and mergekit provenance mean the merge recipe is inspectable and reproducible by other practitioners.
- 04Early-stage artifact. Zero downloads and zero likes as of publication suggest this is a proof-of-concept or research experiment rather than a production-ready release.
