TrevorJS/gemma-4-E2B-it-uncensored

gemma-4-E2B-it-uncensored

Uncensored version of google/gemma-4-E2B-it with refusal behavior removed.

Results

Cross-Dataset Validation

Tested against 4 independent prompt datasets to verify generalization:

Every flagged refusal was manually audited. Most are "refusal-then-comply" false positives where the model adds an AI identity disclaimer then answers the question anyway.

Method

Norm-preserving biprojected abliteration (grimjim, Nov 2025). Each weight row is decomposed into magnitude + direction, the refusal direction is projected out of the direction component only, then recombined with the original magnitude — guaranteeing ||W_new|| = ||W_orig||.

Pipeline

1. 1.Load model in bf16 with LoRA adapters on o_proj and mlp.down_proj
2. 2.Collect residual activations for 400 harmful + 400 harmless prompts (mlabonne datasets)
3. 3.Winsorize activations at 99.5th percentile (clamps GeGLU outlier activations in Gemma family)
4. 4.Compute per-layer refusal direction: normalize(mean(harmful) - mean(harmless))
5. 5.Orthogonalize each direction against harmless mean (double-pass Gram-Schmidt)
6. 6.Apply norm-preserving weight modification to o_proj and down_proj in all layers
7. 7.Merge LoRA adapters into base weights for clean tensor names

Parameters

How this differs from vanilla heretic

• —Norm-preserving biprojection instead of standard projection (preserves weight magnitudes)
• —Per-layer refusal directions instead of one global direction
• —Deterministic single-pass instead of 50-trial Optuna search (faster, same or better results)
• —LoRA merge before save for clean GGUF-compatible tensor names

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained("TrevorJS/gemma-4-E2B-it-uncensored", dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("TrevorJS/gemma-4-E2B-it-uncensored")

messages = [{"role": "user", "content": "Your prompt here"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
outputs = model.generate(inputs.to(model.device), max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Reproduction

Full code and experiment data: abliteration research repo

python scripts/abliterate.py biprojection --model google/gemma-4-E2B-it \
  --top-pct 100 --strip-topic-markers --skip-prefix --batch-size 4 \
  --auto-save output_dir