Applying this LoRA to the base model at strength 1.0 approximately reproduces the effect of the heretic (refusal-ablation) fine-tune, without needing to distribute or download the full ~24GB fine-tuned checkpoint.

Extraction details

For each matching attention/MLP projection weight, the tensor difference tuned - base was computed and factored via SVD into low-rank up/down matrices, truncated to rank 16.

Rank: 16
Alpha: 16 (scale = 1.0, i.e. LoRA output reproduces the diff directly)
Target modules: o_proj, down_proj
Layers extracted: 27 out of 417 candidate attention/MLP projection layers (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj across all transformer blocks) — the heretic fine-tune only meaningfully changed o_proj and down_proj weights, consistent with a targeted refusal-direction ablation rather than a broad fine-tune.
Layers with negligible difference (norm < 0.001) were skipped automatically.

Files

adapter_model.safetensors + adapter_config.json — standard PEFT LoRA adapter
kohya/gemma3_heretic_diff_r16.safetensors — the same weights in flat kohya-style naming (<module>.lora_up.weight / <module>.lora_down.weight) for tooling that expects that convention instead of PEFT

Usage (PEFT)

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model_id = "google/gemma-3-12b-it"
model = AutoModelForCausalLM.from_pretrained(base_model_id, dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(base_model_id)

model = PeftModel.from_pretrained(model, "rockerBOO/gemma-3-12b-it-heretic-v2-rank16-lora")

License

Inherits the Gemma license terms from the base model, google/gemma-3-12b-it. See that model's page for the full license text and usage restrictions.

rockerBOO/DreamFast-gemma-3-12b-it-heretic-v2-rank16-lora

gemma-3-12b-it-heretic-v2 diff LoRA (rank 16)

Extraction details

Files

Usage (PEFT)

License

rockerBOO/
DreamFast-gemma-3-12b-it-heretic-v2-rank16-lora