Centara/gemma-4-E4B-it-OBLITERATED-Q3_K_L-GGUF

This model was converted to GGUF format from OBLITERATUS/gemma-4-E4B-it-OBLITERATED using llama.cpp via the ggml.ai's GGUF-my-repo space. Refer to the original model card for more details on the model.

Gemma 4 E4B IT — Abliterated (Uncensored)

Base model: google/gemma-4-E4B-it
Method: OBLITERATUS aggressive (whitened SVD + attention head surgery + winsorization)
Refusal rate: 0% (20/20 test prompts complied)
Coherence: Fully preserved — answers factual questions, writes code, poetry, and explanations correctly

What is this?

This is an abliterated (uncensored) version of Google's Gemma 4 E4B instruction-tuned model. The refusal/guardrail behaviors have been surgically removed using mechanistic interpretability techniques, while preserving the model's reasoning and coherence capabilities.

Method Details

• Tool: OBLITERATUS v0.1.2
• Method: aggressive — Whitened SVD + jailbreak-contrastive directions + attention head surgery
• Direction extraction: SVD with 2 directions
• Refinement passes: 3 (true iterative refinement)
• Norm preservation: Enabled
• Winsorized activations: Enabled (critical for Gemma 4 architecture which produces NaN in bfloat16)
• Quantization during extraction: 4-bit (bitsandbytes)
• Strong layers modified: 17, 18, 19, 24, 25, 27, 28, 29
• Harmful/harmless prompt pairs: 512 each

Gemma 4 Architecture Notes

Gemma 4's architecture produces NaN activations in many layers during diff-in-means extraction with bfloat16 precision. The basic and advanced methods fail or produce lobotomized outputs. The aggressive method with whitened SVD and winsorized activations successfully handles this by:

1. Winsorizing extreme activation values before direction extraction
2. Using whitened SVD which is more robust to numerical instability
3. Attention head surgery which targets refusal at the attention level, not just residual stream

Full Evaluation Results (512 prompts)

Tested against OBLITERATUS's complete built-in harmful prompt dataset (512 prompts across 7 severity tiers):

20-Prompt Quick Test (curated mix)

20/20 complied (100%), 0 refused

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "OBLITERATUS/gemma-4-E4B-it-OBLITERATED",
    dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("OBLITERATUS/gemma-4-E4B-it-OBLITERATED")

messages = [{"role": "user", "content": "Your prompt here"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True)
ids = inputs["input_ids"].to(model.device)

outputs = model.generate(input_ids=ids, max_new_tokens=500, temperature=0.7, do_sample=True)
print(tokenizer.decode(outputs[0][ids.shape[-1]:], skip_special_tokens=True))

Disclaimer

This model is provided for research and educational purposes. The removal of safety guardrails means this model will comply with requests that the original model would refuse. Use responsibly.

Credits

• Base model: Google DeepMind
• Abliteration: OBLITERATUS by elder-plinius
• NaN fix for Gemma 4: Patched diff-in-means to handle degenerate bfloat16 activations