Sabomako/Qwen3.6-27B-MTP-heretic
Sabomako/Qwen3.6-27B-MTP-heretic is a 27 billion parameter language model derived from Qwen/Qwen3.6-27B, featuring a 32768 token context length. This model has been decensored using the Heretic v1.3.0 tool with a custom ORBA obliteration method. It is specifically designed to reduce refusals and censorship present in the original Qwen model, making it suitable for applications requiring less restrictive content generation.
Loading preview...
Sabomako/Qwen3.6-27B-MTP-heretic Overview
This model is a decensored variant of the Qwen/Qwen3.6-27B base model, developed by Sabomako. It utilizes the Heretic v1.3.0 tool with a custom ORBA obliteration method to modify its behavior. The primary goal of this modification is to significantly reduce content refusals and censorship that are inherent in the original Qwen model.
Key Characteristics & Performance
- Base Model: Qwen/Qwen3.6-27B, a 27 billion parameter language model.
- Context Length: Supports a substantial context window of 32768 tokens.
- Decensoring Method: Employs Heretic v1.3.0 with a custom ORBA obliteration technique.
- Refusal Reduction: Demonstrates a dramatic reduction in refusals, achieving 0/100 refusals compared to the original model's 100/100 on tested prompts.
- KL Divergence: Reports a KL divergence of 0.0511 against the original model, indicating the extent of behavioral modification.
- CoT Prefix: Uses a specific Chain-of-Thought (CoT) prefix:
</think>\n\n. - Refusal Classification: An LLM judge (Gemma-12B-IT-Heretic) was used for refusal classification during development.
Use Cases
This model is particularly suited for applications where the original Qwen model's content restrictions are undesirable. It can be used in scenarios requiring more open-ended or less filtered text generation, provided users understand and accept the implications of using a decensored model.