MuXodious/Maginum-Cydoms-24B-absolute-heresy
MuXodious/Maginum-Cydoms-24B-absolute-heresy is a 24 billion parameter language model, fine-tuned from a Maginum-Cydoms-24B base using P-E-W's Heretic abliteration engine. This model is a complex merge of several 24B models, including Mistral-Small-3.2-24B-Instruct, Magidonia-24B, Precog-24B, and Cydonia-24B, utilizing TIES, DELLA, and SLERP merge methods. It is specifically engineered to exhibit an "Absolute Heresy" classification, characterized by a low refusal rate (4/100) and low KL Divergence (0.0307), indicating a significant deviation from its base models' original doctrines.
Loading preview...
Maginum-Cydoms-24B-absolute-heresy: A Deep Dive
This model, developed by MuXodious, is a 24 billion parameter variant of the Maginum-Cydoms-24B family, distinguished by its unique "heretication" process. It was created using P-E-W's Heretic (v1.1.0) abliteration engine, which applies a Magnitude-Preserving Orthogonal Ablation PR to modify the model's behavior.
Key Characteristics & Heretication
The model achieves an "Absolute Heresy" classification, defined by a very low refusal rate of 4/100 and a KL Divergence of 0.0307. This indicates a significant alteration from its original training, aiming to reduce refusals and potentially bypass certain safety alignments present in its base models. The Heresy Index, while arbitrarily inspired by Warhammer 40K, provides a metric for this behavioral shift, with "Absolute Heresy" signifying less than 10/100 refusals and a KL Divergence below 0.10.
Merge Details
Maginum-Cydoms-24B-absolute-heresy is a sophisticated merge of multiple pre-trained language models, orchestrated using mergekit. The merge process involved several advanced methods:
- TIES: A technique for merging models by identifying and combining their important parameters.
- DELLA: Another merging approach that contributes to the model's unique characteristics.
- SLERP: Spherical Linear Interpolation, used for smoothly blending the parameters of the merged models.
The base models contributing to this merge include anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only, TheDrummer/Magidonia-24B-v4.3, TheDrummer/Precog-24B-v1, zerofata/MS3.2-PaintedFantasy-v3-24B, TheDrummer/Cydonia-24B-v4.3, ReadyArt/4.2.0-Broken-Tutu-24b, and zerofata/MS3.2-PaintedFantasy-v2-24B.