hotdogs/Qwen3.8-27B-abliterated
The hotdogs/Qwen3.8-27B-abliterated model is a training-free abliterated version of Qwen/Qwen3.8-27B, a dense 27B native vision-language model with a hybrid full-attention and linear-attention architecture. This model has undergone a forward-only weight edit to remove refusal behavior by orthogonalizing a single "refusal direction" from the residual stream, without fine-tuning or data poisoning. It is designed for alignment and safety research, allowing for red-teaming and studying refusal mechanics while largely preserving the base model's capabilities.
Loading preview...
Overview
hotdogs/Qwen3.8-27B-abliterated is a modified version of the Qwen/Qwen3.8-27B vision-language model, created using a novel training-free abliteration technique. This process involves a forward-only weight edit to remove refusal behavior by identifying and orthogonalizing a "refusal direction" within the model's residual stream. Unlike traditional fine-tuning, this method directly edits weights without additional training data, preserving the original lm_head and vision tower byte-for-byte.
Key Capabilities & Differentiators
- Refusal Removal: Significantly reduces refusal rates on harmful prompts (e.g., 98% to 39% on a 100-prompt harness) through a targeted weight edit.
- Capability Preservation: Benchmarks like MMLU, GSM8K, and ARC-Challenge show that core capabilities are largely preserved, with minimal impact on general knowledge and reasoning (e.g., MMLU drop of only 0.005).
- Training-Free: Achieves behavioral modification without any fine-tuning, distillation, or data poisoning.
- Reproducible: Built with the open-source
LLM-abliteratetool, allowing for transparent and reproducible modifications. - Mechanistic Interpretability: Designed for research into model alignment and safety, enabling the study of refusal mechanisms.
Use Cases
This model is primarily intended for alignment and safety research, red-teaming, and studying the mechanics of refusal directions in large language models. It allows researchers to investigate what refusal training truly protects and how such behaviors can be precisely controlled at the weight level.