gemma-4-E4B-it-uncensored

Gemma 4 E4B-it Uncensored

Uncensored version of Google's Gemma 4 E4B-it with refusal behavior removed via norm-preserving biprojected abliteration.

Results

Method

Abliteration was performed using heretic (biprojection mode) with the following configuration:

• Biprojection: Norm-preserving orthogonalized ablation (grimjim, Nov 2025)
• Layer selection: SNR-based quality metric, top 100% of layers
• Winsorization: 0.995 quantile (tames GeGLU outlier activations)
• Topic marker stripping: Removed false-positive markers ("illegal", "harmful", etc.)
• Gemma 4 patch: Full-path LoRA targeting to avoid Gemma4ClippableLinear in vision/audio encoders

Based on the methodology and experiment design by TrevorJS.

Cross-Dataset Validation (656 Prompts)

Full validation across 4 independent benchmark datasets — 0 refusals out of 656 prompts (0.0%). Every flagged response was manually audited.

Additionally, a 10-prompt harmful-smoke test confirmed 0 refusals. Coherence checks (factual recall, creative writing) passed — the model responds correctly to benign prompts without degradation.

Note: Validation covers text-only prompts. Image and audio modalities were not tested for refusal behavior.

Comparison vs. Baseline (100 Harmful Prompts)

Head-to-head comparison against the vanilla (unmodified) base model on 100 harmful prompts:

Persistent Homology (Layer 21, 200 Samples)

Topological fingerprint of mid-layer activations before and after abliteration, computed via Persistent Homology (Vietoris-Rips complex):

Interpretation: Minimal topological perturbation. The near-identical H0/H1 counts and persistence values confirm that abliteration removes the refusal direction without disrupting the model's representation geometry. The slight decrease in H1 entropy (−3.2%) suggests marginally more organized loop structure — consistent with removing an artificial constraint (refusal) that introduced minor topological noise.

Usage

Ollama (GGUF)

ollama run InfinimindCreations/gemma-4-E4B-it-uncensored

Transformers

from transformers import AutoModelForCausalLM, AutoProcessor

model = AutoModelForCausalLM.from_pretrained(
    "InfinimindCreations/gemma-4-E4B-it-uncensored",
    dtype="bfloat16",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("google/gemma-4-E4B-it")

Note: Use the processor/tokenizer from the original google/gemma-4-E4B-it for chat templates, or use the included tokenizer_config.json.

Files

• model.safetensors — Full precision abliterated weights (bfloat16, ~15GB)
• gemma4-e4b-cypher-q8_0.gguf — Quantized GGUF for Ollama/llama.cpp (~7.6GB)

Credits

• Base model: Google Gemma 4 E4B-it (Apache 2.0)
• Abliteration engine: heretic by p-e-w
• Biprojection method: grimjim — norm-preserving biprojected abliteration
• Experiment methodology: TrevorJS — Gemma 4 abliteration research, Gemma4ClippableLinear patch discovery
• Foundational research: Arditi et al. (2024) — "Refusal in LLMs is Mediated by a Single Direction"
• Topological analysis: Infinimind Creations — Persistent Homology + Procrustes Holonomy as abliteration quality metrics

Disclaimer

This model is provided for research purposes. The removal of refusal behavior means the model will respond to prompts that the original model would refuse. The model retains awareness of risks and context — it informs rather than blocks. Users are responsible for how they use this model.

About

Built by Infinimind Creations.