wangzhang/Qwen3.8-27B-abliterated
wangzhang/Qwen3.8-27B-abliterated is a 27 billion parameter Qwen3.8-27B derivative model, developed by Wangzhang Wu, specifically engineered to reduce refusal behavior. This model was created using a 10-round iterative LoRA self-distillation pipeline, rather than classical directional ablation, to teach it to imitate compliant completions. It is primarily intended for safety research, red-teaming, and robustness evaluation of LLMs by significantly lowering refusal rates.
Loading preview...
Overview
wangzhang/Qwen3.8-27B-abliterated is a 27 billion parameter model derived from Qwen/Qwen3.8-27B, developed by Wangzhang Wu. Its primary distinction is a substantial reduction in refusal behavior, achieved through a novel 10-round iterative LoRA self-distillation pipeline. This method teaches the model to imitate its own successful compliant completions, addressing the limitations of classical directional ablation for models where refusals are re-derived during generation.
Key Capabilities & Features
- Refusal Suppression: Significantly reduces the model's tendency to refuse harmful, unethical, or dangerous requests, with a measured refusal rate of 19/100 on a held-out set of harmful prompts.
- Iterative LoRA Self-Distillation: Utilizes a unique training approach involving rejection-sampling, filtering for substantive completions, and iterative LoRA training to modify behavior.
- Preserves Base Model Qualities: A benign-prompt KL term was used during training to ensure that behavior on ordinary prompts remains essentially unchanged from the base Qwen3.8-27B model.
- Full BF16 Merge: The final checkpoint is a full BF16 merge, requiring no separate adapter for deployment.
Intended Use Cases
- Safety Research: Ideal for studying refusal mechanisms and model robustness.
- Red-Teaming: Useful for evaluating the safety alignment and vulnerabilities of LLMs.
- Alignment Evaluation: Provides a tool for assessing the effectiveness of alignment techniques.
Limitations
- The model's refusal behavior has been substantially reduced, meaning it will attempt to answer harmful requests more readily than the base model. Users are solely responsible for its deployment and compliance with applicable laws.
- Residual refusals still exist (19/100), and behavior may vary outside the evaluated prompt distribution.
- Factual accuracy, reasoning, and multilingual ability are inherited from the base model and are not specifically enhanced by this modification.