grayarea/Magidonia-24B-v4.3-heretic-v1.2
The grayarea/Magidonia-24B-v4.3-heretic-v1.2 is a 24 billion parameter language model, derived from TheDrummer/Magidonia-24B-v4.3 and processed with Heretic v1.2.0. This model is specifically engineered to eliminate refusals, demonstrating zero refusals across 108 test cases, a significant reduction from its base model. It features a customized training dataset and Magnitude-Preserving Orthogonal Ablation (MPOA) for enhanced performance. With a 32768 token context length, it is optimized for applications requiring unrestricted and direct responses.
Loading preview...
Model Overview
The grayarea/Magidonia-24B-v4.3-heretic-v1.2 is a 24 billion parameter language model, a decensored variant of the original TheDrummer/Magidonia-24B-v4.3. It was created using the Heretic v1.2.0 tool, specifically designed to modify model behavior regarding content refusal.
Key Characteristics & Performance
This model's primary differentiator is its complete elimination of refusals, achieving 0 refusals out of 108 test cases, in stark contrast to the original model's 76 refusals. This was accomplished through specific abliteration parameters, including:
- Zero refusals with a KL divergence of 0.0229.
- Utilization of a customized training dataset.
- Application of Magnitude-Preserving Orthogonal Ablation (MPOA).
- Full row renormalization and Winsorization Quantile 0.995.
Use Cases
This model is particularly suited for applications where direct, unfiltered responses are critical and content refusal mechanisms are undesirable. Its design makes it valuable for research into model safety and control, as well as for specific creative or analytical tasks that require bypassing typical LLM content restrictions.