This is a merge of pre-trained language models created using mergekit.

Experimental component for a project. Extremely dumb config that I expect to have to tune. Mostly better than R1, less hallucination in vision tasks. Doing TA on the abliterated base seems to restore the refusal vector somewhat. Experimentation is being done to try and transplant abliteration directly, ahead of a Mergekit PR to fix LoRA extraction on Gemma3. Tiger is the domineering "flavor"; could be toned down.

This model was merged using the Task Arithmetic merge method using grimjim/gemma-3-12b-it-norm-preserved-biprojected-abliterated as a base.

The following models were included in the merge:

The following YAML configuration was used to produce this model: