MuXodious/QuasiStarSynth-12B-noslop-absolute-heresy
MuXodious/QuasiStarSynth-12B-noslop-absolute-heresy is a 12 billion parameter language model, fine-tuned from QuasiStarSynth-12B-noslop using P-E-W's Heretic v1.1.0 ablation engine with Magnitude-Preserving Orthogonal Ablation. This model achieves an "Absolute Heresy" classification with 6/100 refusals and 0.0080 KL Divergence, indicating a significant reduction in refusal behavior. It is intended for further training with high-quality prose datasets or as a base for model mergers to enhance originality and reduce clichés in creative writing and roleplay scenarios.
Loading preview...
Model Overview
MuXodious/QuasiStarSynth-12B-noslop-absolute-heresy is a 12 billion parameter model derived from the QuasiStarSynth-12B-noslop base. It was processed using P-E-W's Heretic (v1.1.0) ablation engine, specifically with Magnitude-Preserving Orthogonal Ablation enabled. This process aims to modify specific model behaviors, particularly reducing refusals and "slop" in its output.
Heretication and Noslopfication
The model has undergone a "heretication" process, resulting in an Absolute Heresy classification. This is characterized by a low refusal rate (6/100) and a KL Divergence of 0.0080, indicating a significant shift from its initial refusal behavior (92/100). The "noslopfication" process, which removed "slop direction" from the base model, resulted in a slop score of 39/100 from an initial 89/100.
Intended Use and Hypothesis
While the removal of "slop direction" is hypothesized to improve originality, reduce clichés, and lower redundancy in creative writing and roleplay, this model should be further trained with high-quality prose datasets or used as a base in mergers to fully realize these enhancements. The base QuasiStarSynth-12B model supports ChatML style messages and is stable at 12k tokens, with potential for extended contexts.