Ilya626/gemma-4-12B-it-SDFT-Heretic-LoRA

Gemma 4 12B SDFT Heretic LoRA

Experimental adapters. Safety guardrails may be weakened or compromised; outputs may be unsafe, illegal, offensive, or incorrect. Review outputs and do not use for safety-critical or unsupervised deployment.

Files

• raw_r6/: final trained PEFT LoRA and its Q8_0 llama.cpp LoRA (r=6, lora_alpha=9, rsLoRA).
• smoothed_r6/: rank-6 PEFT approximation and its Q8_0 llama.cpp LoRA (r=6, lora_alpha=6, standard LoRA).

Scaling

Use external adapter scale 1.0 for every provided adapter.

PEFT already applies the raw adapter's internal rsLoRA scaling (9/sqrt(6)). The raw GGUF stores the equivalent converted alpha, so do not multiply either form again. The smoothed adapters have the intended delta baked into their factors. For PEFT merging, load normally and call merge_and_unload(). For llama.cpp, attach the selected GGUF LoRA at scale 1.0.