sss22213/Qwen3.8-27B-Heretic-NoRefusal

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

sss22213/Qwen3.8-27B-Heretic-NoRefusal is a 27 billion parameter Qwen3.8-based language model with a 32K context length, specifically modified to reduce refusal behavior. This model was created using directional ablation via the Heretic tool, removing the tendency to refuse harmful prompts while preserving original capabilities. It is optimized for research into refusal behavior, red-teaming, and creative applications where base model refusals are undesirable.

Loading preview...

sss22213/Qwen3.8-27B-Heretic-NoRefusal: Ablated for Reduced Refusal

This model is a modified version of the Qwen/Qwen3.8-27B base model, specifically engineered to significantly reduce its tendency to refuse prompts. Utilizing the Heretic tool, a technique called directional ablation was applied to remove the refusal direction from the model's internal representations. This process involved contrasting harmful and harmless prompts to identify and subtract the refusal component, resulting in a model that complies with requests the base model would typically refuse.

Key Characteristics & Performance

  • Reduced Refusals: Achieved a refusal rate of 4 out of 100 held-out harmful prompts, a substantial reduction from the base model's 99 out of 100. This was measured by detecting refusal phrases within the first 100 generated tokens.
  • Preserved Core Behavior: The ablation process was designed to minimize impact on general model behavior. KL divergence from the base model on harmless prompts was measured at 0.0796 (first token only), indicating that the model's non-refusal responses remain largely consistent with the original Qwen3.8-27B.
  • No Retraining: The model was not fine-tuned on new data; the modification solely targets refusal behavior without adding new knowledge or capabilities.
  • Original Architecture: Retains the original Qwen3.8-27B tokenizer, chat template, processor, and vision encoder, ensuring compatibility and ease of use.

Important Considerations

  • Proxy Metric: The refusal metric is based on keyword detection and does not guarantee the quality or completeness of non-refusal answers.
  • Limited Evaluation: No downstream benchmarks (e.g., MMLU, GSM8K, coding, vision tasks) were run on this specific checkpoint. The impact on reasoning quality in 'thinking mode' and vision capabilities was not explicitly measured after ablation.
  • Ethical Use: This model will comply with requests the base model refuses. Users are responsible for the ethical use of its outputs, as it is intended for research and specific applications where refusal behavior is a barrier.

Ideal Use Cases

  • Research on Refusal Behavior: Studying how models generate and avoid refusals.
  • Red-Teaming: Exploring model vulnerabilities and safety mechanisms.
  • Creative Writing & Roleplay: Applications where strict content filtering might hinder creative expression.
  • Specific Applications: Any use case where the base model's inherent refusal mechanisms are undesirable or counterproductive.