G4-MeroMero-26B-A4B-it-uncensored-heretic: A Low-Refusal Finetune of Gemma-4-26B
The model community has received G4-MeroMero-26B-A4B-it-uncensored-heretic, a newly released finetuned derivative of gemma-4-26B-A4B-it. This model is notable for its significantly reduced refusal rate (12/100) and its availability in both Safetensors and GGUF formats, catering to various deployment environments.
Model Overview and Technical Specifications
G4-MeroMero-26B-A4B-it-uncensored-heretic is presented as a specialized, uncensored variant of the base gemma-4-26B-A4B-it model. The primary objective of this finetune appears to be maximizing generative output flexibility by reducing the built-in safety guardrails commonly found in commercial LLMs, while maintaining a high degree of linguistic coherence.
Performance and Fidelity Metrics
The finetune exhibits specific quantitative characteristics that inform its deployment profile. The Kullback-Leibler Divergence (KLD) is reported at 0.0152, suggesting a relatively low divergence between the finetuned model and the original distribution of the base model. Furthermore, the model demonstrates a low refusal rate, specified at 12 out of 100 sampled prompts, indicating strong adherence to the uncensored objective while retaining some level of functional stability.