Gemma 4 E4B for ComfyUI

ComfyUI-format LoRAs for the Gemma 4 E4B text encoder.

Use

LoRAs into models/loras/, plus gemma4_e4b_it_bf16.safetensors from Comfy-Org/gemma-4 into models/text_encoders/.

CLIPLoader -> LoRA on CLIP (strength 1.0) -> Generate Text

• Leave CLIPLoader's type on the default — Gemma 4 E4B is detected from the weights.
• Strength 1.0 is the full OBLITERATED model; lower values blend back toward stock behaviour.
• Use greedy decoding (sampling_mode: off). The checks below were run with thinking off.
• This is not a refusal-only edit. OBLITERATED also shifts ordinary answers (stock: "The capital of France is Paris." → OBLITERATED: "Here is your answer: Paris"), and the LoRA carries that too. Expect conditioning for benign prompts to differ from the stock encoder.

How it was made

A plain weight diff, OBLITERATED − gemma4_e4b_it_bf16, factored per matrix with SVD. Unlike a pure single-direction abliteration the edit isn't exactly low rank — OBLITERATUS's aggressive mode takes two directions plus a norm-preserving rescale, which leaves a long tail — so each matrix keeps the smallest rank reaching 99.9% of its diff energy, capped at 64. Most layers need rank 2–4; the capped ones still keep ≥99.74%.

• 137 matrices, language model only: embed_tokens and every attention (q/k/v/o_proj) and MLP (gate/up/down_proj) projection in layers 17–40. The vision and audio towers, norms and the per-layer embedding table are untouched (diffs ≤3e-8, fp16 round-trip noise only).
• Checked inside ComfyUI against the OBLITERATED weights loaded into the same encoder: next-token KL to OBLITERATED goes from 3.25 / 6.17 (stock; refusal / benign prompts) to 0.0009 / 0.0010 (stock + LoRA), and greedy outputs match it token-for-token on all 8 test prompts.

Keys are ComfyUI's generic text-encoder form (text_encoders.gemma4.transformer.model.<module>.lora_{down,up}.weight), bf16, with alpha = rank (scale 1.0).

Credits

Gemma 4 E4B (Apache-2.0) · OBLITERATUS (Apache-2.0) — all uncensoring behaviour is theirs, this only re-expresses it as a LoRA · Comfy-Org for the ComfyUI-format encoder the diff was taken against.