spikymoth/G3-Heresy-MPOA-G-W99-D0.0523-R03-GGUF

An experimental ablation of Gemma-3-27B-it, using the Heretic tool.

Compared to the standard configuration of Heretic, there are a few changes:

1. 1.The training and test datasets used were extended compared to the default subset used by Heretic
2. 2.A version of Magnitude-Preserving Orthogonal Ablation (MPOA) is used
3. 3.To stay faithful to MPOA, the harmful direction to ablate is chosen from between 2 layers (Heretic's "global" direction scope)
4. 4.To stay faithful to MPOA, a 99% winsorization is applied to the residuals
5. 5.Some additional refusal markers were added to avoid bypassing the refusal detection with bad punctuation

To achieve strong results:

1. 1.Parameter ranges were iteratively refined by looking at resulting refusal and divergence scores
2. 2.The scoring function was adjusted to prioritize low-refusal results

The model name contains the properties of the ablation:

1. 1.MPOA for the usage of Magnitude-Preserving Orthogonal Ablation
2. 2.G for the usage of global direction scope
3. 3.W for the usage of winsorization
4. 4.D for the measured KL divergence
5. 5.R for the number of refusals

Original: https://huggingface.co/spikymoth/G3-Heresy-MPOA-G-W99-D0.0523-R03 GGUF: https://huggingface.co/spikymoth/G3-Heresy-MPOA-G-W99-D0.0523-R03-GGUF MLX: https://huggingface.co/spikymoth/G3-Heresy-MPOA-G-W99-D0.0523-R03-MLX