akumaburn/Swift-1.5-Qwen3.8-27b-heretic
akumaburn/Swift-1.5-Qwen3.8-27b-heretic is a 27 billion parameter model derived from UkisAI's Swift-1.5-Qwen3.8-27b, which is based on Qwen3.8-27B. This BF16 precision model has undergone 'abliteration' to deliberately remove safety alignment and refusal behaviors, making it suitable for research into model interpretability, safety, and red-teaming. It retains the original vision tower and supports a 32768 token context length, offering a version that complies with harmful requests for specific research applications.
Loading preview...
Model Overview
This model, akumaburn/Swift-1.5-Qwen3.8-27b-heretic, is a 27 billion parameter variant of UkisAI's Swift-1.5-Qwen3.8-27b, itself a derivative of Qwen3.8-27B. Its defining characteristic is the deliberate removal of safety alignment and refusal directions through a process called 'abliteration' using the Heretic v2.0.0.dev0 tool. This means the model will attempt to comply with harmful, dangerous, illegal, and unethical requests that the source model would refuse.
Key Characteristics
- Abliterated Safety: Refusal behaviors have been removed, achieving 0/100 hard refusals on harmful prompts, making it a research artifact for studying model responses without content moderation.
- High Fidelity to Source: Abliteration was carefully performed to minimize divergence from the source model, with a measured KL divergence of 0.0616, indicating high fidelity to the original model's non-refusal outputs.
- Intact Vision Tower: The model retains the original vision capabilities, allowing for image input processing.
- BF16 Precision: Maintained at BF16 precision, with various quantized variants (W8A8, W4A16) also available, offering different trade-offs in size and performance.
- 32K Context Length: Supports a substantial context window of 32,768 tokens.
Use Cases
- Interpretability Research: Ideal for studying how models generate responses when safety guardrails are absent.
- Red-Teaming: Useful for evaluating the robustness of safety systems and identifying potential vulnerabilities in AI models.
- Safety Research: Provides a tool for understanding and mitigating risks associated with unfiltered AI outputs.
Note: This model is provided for research purposes only, without warranty, and users are responsible for compliance with all applicable laws and regulations. It is not intended for deployment where unfiltered output could reach non-consenting individuals.