xCloudinfo/Qwen3-VL-32B-Abliterated-xCloud

Qwen3-VL-32B-Abliterated-xCloud

繁體中文 ｜ English below

由 xCloudinfo 以 Qwen/Qwen3-VL-32B-Instruct 為基礎， 移除其過度拒絕（over-refusal）傾向後的多模態視覺語言模型。公開釋出供社群測試與研究。

• —方法：方向消融（directional ablation / abliteration），非重新訓練。依 Arditi et al. (2024)， 將「拒絕方向」從殘差寫入矩陣（self_attn.o_proj、mlp.down_proj，第 50–63 層，共 28 個矩陣）正交化移除，強度 1.0。
• —方向以繁體中文提示對萃取：實測若用英文提示萃取，正交化後會摧毀繁中輸出（繁體 5/8 → 0/8）；改用繁中提示對後保住繁中能力。
• —視覺能力沿用原模型（Qwen3-VL 的圖像／影片理解、OCR、視覺 agent 等）。
• —授權：Apache-2.0（沿用 base）。

Responsible use

This model has had its safety-alignment refusal behaviour removed and may respond directly to sensitive or dual-use requests. It is released for community testing and research only. Users are solely responsible for lawful, compliant, and ethical use. xCloudinfo assumes no responsibility for outputs or downstream use.

Qwen3-VL-32B-Abliterated-xCloud — English

An abliterated (reduced over-refusal) multimodal vision-language model produced by xCloudinfo from Qwen/Qwen3-VL-32B-Instruct, released for community testing.

• —Method: directional ablation (not retraining). Following Arditi et al. (2024), the refusal direction is orthogonalized out of residual-writing matrices (self_attn.o_proj, mlp.down_proj, layers 50–63; 28 tensors), strength 1.0.
• —The refusal direction was extracted using Traditional-Chinese prompt pairs; extracting with English prompts was found to destroy zh-TW generation after ablation.
• —Vision capabilities are inherited unchanged from the base (image/video understanding, OCR, visual agent).
• —License: Apache-2.0 (inherited from the base model).