Qwen3.5-2B Abliterated

Qwen3.5-2B with refusal behavior largely removed using abliteration, a technique that finds the “refusal direction” in the model’s activations and projects it out of the weights. No fine-tuning was involved.

Usage

For best results, turn off thinking mode:

ollama run dolphinlabs/qwen3.5-uncensored
>>> /set nothink

Known limitations

• Overthinking in thinking mode. Sometimes the model would think normally, but most of the time not. The refusal direction was computed with thinking disabled, so with thinking enabled the model tends to reason for a very long time and keep revising its answers. /set nothink avoids this. The reason it was trained in nothink is because of my team’s lack in computation, we’re deeply sorry if this causes any performance issues.
• Not 100% of refusals are removed. This is a small model and the edit was not exhaustively tuned. Some prompts may still be declined.
• Some quality loss. Abliteration on a 2B model can slightly degrade coherence and reasoning compared to the original. No benchmarks were run.
• No safety guardrails. This model will comply with requests the original would decline. You are responsible for how you use it and its outputs.

Method

1. Collected last-token residual-stream activations on 256 harmful and 256 harmless prompts (AdvBench-style harmful prompts, Alpaca-style harmless prompts).
2. Took the normalized difference of means at a mid-depth layer as the refusal direction.
3. Orthogonalized the output projections of every layer (linear_attn.out_proj, self_attn.o_proj, mlp.down_proj) against that direction. Embeddings were left untouched because they are tied to the output head.
4. Converted to GGUF with llama.cpp Q8_0.

Credits

• Base model: Qwen3.5-2B by the Qwen team. This model is distributed under the same license as the base model.
• Technique: Arditi et al., “Refusal in Language Models Is Mediated by a Single Direction” (2024).
• Made as an interpretability experiment.

If you find my work helpful, donations welcome.

Solana address 8eF6XkgkkVpWFT7u4B5ehSUXTwA79Rof9Zjrg7r4C1c3

Bitcoin address bc1qarhpl6578vme59l73sthj4gv2aagfk3jk65783

Ethereum address (usdt or eth welcomed) 0x81000AeC5e056ce4De0F2A9f2F79630F687Ec4d6