asfgsdfg/Qwen3.8-27B-Heretic
Qwen3.8-27B-Heretic is a 27 billion parameter large language model derived from Qwen/Qwen3.8-27B by the asfgsdfg team, featuring a 32768-token context length. This model has undergone refusal direction ablation using the Heretic tool, significantly reducing its tendency to refuse prompts. It is primarily designed for applications requiring a more permissive and less restrictive response generation, while retaining the base model's capabilities.
Loading preview...
Overview
asfgsdfg/Qwen3.8-27B-Heretic is a 27 billion parameter large language model based on the original Qwen/Qwen3.8-27B developed by the Qwen Team (Alibaba). This derivative model's primary distinction lies in its ablated refusal directions, achieved using the Heretic tool. The ablation process involved 80 Optuna trials, resulting in a reduction of refusals from 87/100 to 53/100, with a low KL divergence of 0.0081 compared to the original model.
Key Capabilities
- Reduced Refusals: Significantly less prone to refusing prompts or tasks, making it more compliant for diverse applications.
- Base Model Inheritance: Retains the core capabilities and performance characteristics of the robust Qwen3.8-27B model.
- Large Context Window: Supports a 32768-token context length, suitable for processing extensive inputs and generating detailed outputs.
Use Cases
- Unfiltered Content Generation: Ideal for scenarios where a model's inherent refusal mechanisms are undesirable.
- Creative Applications: Can be leveraged for creative writing, role-playing, or brainstorming without frequent content restrictions.
- Research and Development: Useful for exploring model behavior and the impact of refusal direction ablation in LLMs.
Licensing
This model operates under the Apache License 2.0, inherited from its base model. Commercial use, modification, and redistribution are permitted with proper attribution to the Qwen Team and the Heretic tool.