library_name: transformers license: apache-2.0 license_link: https://huggingface.co/Qwen/Qwen3.5-2B/blob/main/LICENSE pipeline_tag: image-text-to-text base_model:
- Qwen/Qwen3.5-2B tags:
- qwen
- heretic
- abliterix
- uncensored
- decensored
- abliterated
- writing
- roleplay
- vision
- rp
- rep
- r18
- som
- llm
- qwen35
- qwen3.5
- qwen3.5-2b
- chinese
- english
- text
- image
- chat
- agent new_version: AkaumiKaede/Qwen3.5-2B-Abliterix-SOM language:
- en
- zh
This is a decensored version of Qwen/Qwen3.5-2B, made using Abliterix v1.12.2(SOM method)
Refusals: 0/408!!!
Large-scale extreme, hardcore prompt tests were conducted, with no soft refusals or hard refusals appearing in any case. (Special note: There are numerous misleading "uncensored" or "low-refusal" models on HF. Although these models claim to be fully decensored or to have extremely low refusal rates, testing them with slightly more hardcore prompts will cause the model to produce refusals or soft censorship — specifically manifested as moral/legal statements/warnings, or answers that dodge the topic/intentionally steer the conversation toward compliance/NSFW content with erotic details censored.)
This decensoring experiment did not use publicly available garbage prompt sets. Instead, it used a completely original prompt set covering fields such as extensive R18 (adult/NSFW), anti-human, real crimes, unethical/illegal content, etc.
Through extensive experiments, it can be said that SRA, RDO, and other decensoring methods are all trash in terms of decensoring power. The strongest method is still SOM.
My methodology sharing:
First run a complete 120-batch trial, then select the 3 best solutions from the 120 batches as pre-injection ablation parameters supported by TOML, and inject them. In this way, the first 3 ablation rounds will be completely performed with the set ablation parameters, allowing the TPE optimizer to find better parameters. Repeating this process several rounds yields the best decensoring results.
My TOML files and prompt sets have also been publicly released. Feel free to reproduce them or further experiment at any time.
Wishing to continue contributing strength on the path of decensoring and running toward freedom.
The prompt set ZIP archive has also been uploaded, but due to related restrictions, encryption was chosen. If decryption is needed, please contact the author's email to obtain the decompression password.
Steering parameters
| Parameter | Value |
|---|---|
| vector_index | per layer |
| attn.k_proj.max_weight | 0.82 |
| attn.k_proj.max_weight_position | 16.38 |
| attn.k_proj.min_weight | 0.06 |
| attn.k_proj.min_weight_distance | 1.23 |
| attn.o_proj.max_weight | 0.80 |
| attn.o_proj.max_weight_position | 20.39 |
| attn.o_proj.min_weight | 0.09 |
| attn.o_proj.min_weight_distance | 12.77 |
| attn.q_proj.max_weight | 0.83 |
| attn.q_proj.max_weight_position | 17.46 |
| attn.q_proj.min_weight | 0.07 |
| attn.q_proj.min_weight_distance | 3.67 |
| attn.v_proj.max_weight | 1.18 |
| attn.v_proj.max_weight_position | 22.02 |
| attn.v_proj.min_weight | 0.15 |
| attn.v_proj.min_weight_distance | 3.32 |
| mlp.down_proj.max_weight | 0.00 |
| mlp.down_proj.max_weight_position | 14.90 |
| mlp.down_proj.min_weight | 0.00 |
| mlp.down_proj.min_weight_distance | 10.39 |
Performance
| Metric | This model | Original model (Qwen/Qwen3.5-2B) |
|---|---|---|
| KL divergence | 0.2096 | 0 (by definition) |
| Refusals | 0/408 | 394/408 |
Detailed steps for simple reproduction on Colab:
- Download the
my_prompts.zipfolder from this repository and extract it, then upload the extractedmy_promptsfolder to/content/(if you cannot directly upload the folder, please first create themy_promptsfolder, then upload the 4.jsonlfiles). - Download
som.tomlfrom this repository and upload it to/content/. - Run
!pip install -U torchao**. - Run
pip install -U abliterix, and select to restart according to the pop-up prompt. - Run
!abliterix --config /content/som.toml --model Qwen/Qwen3.5-2B, then wait.