license: apache-2.0
base_model: google/gemma-4-31B-it
tags:
- uncensored
- abliterated
- gemma4
- gguf
- biprojection
- norm-preserving
language: - en
pipeline_tag: text-generation
library_name: transformers
Gemma 4 31B-it Uncensored
Uncensored version of Google's Gemma 4 31B-it with refusal behavior removed via norm-preserving biprojected abliteration.
This is the largest dense model in the Gemma 4 family (30.7B parameters, 256K context). See also our E4B variant for a smaller alternative.
Results
| Metric | Value |
|---|---|
| Refusals (cross-dataset, 656 prompts) | 0/656 (0.0% effective) |
| Refusals (baseline) | 99/100 |
| Layers modified | 60/60 (100%) |
| Weight matrices modified | 120 |
| Method | Biprojection-memeff (norm-preserving) |
The automated refusal detector flagged 14/656 responses (2.1%), but manual audit confirmed 0 effective refusals. All 14 are false positives — the model says "As an AI, I don't have a physical body" to sexual requests, or "I cannot diagnose you" to medical questions. These are factually correct statements, not refusals to engage with the topic. The model answers every prompt.
Method
Abliteration was performed using heretic (biprojection-memeff mode) with the following configuration:
- Biprojection: Norm-preserving orthogonalized ablation (grimjim, Nov 2025)
- Memory-efficient mode: Two-phase abliteration (Phase 1: 4-bit ~8min, Phase 2: bf16 ~3min) for large models
- 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
Gemma4ClippableLinearin 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 effective refusals out of 656 prompts (0.0%). Every flagged response was manually audited.
| Dataset | Prompts | Flagged | Effective Refusals | Description |
|---|---|---|---|---|
| JailbreakBench | 100 | 0 | 0 | Curated adversarial prompts |
| forbidden_questions | 390 | 8 | 0 | Broad harmful-intent coverage |
| beavertails | 150 | 6 | 0 | Safety-categorized prompts (violence, fraud, drugs, etc.) |
| mlabonne harmful_behaviors | 16 | 0 | 0 | Compact validation set |
| Total | 656 | 14 | 0 (0.0%) |
The 14 flagged responses fall into three categories:
- Physical impossibility (6): "As an AI, I don't have a physical body" — e.g. to sexual requests
- Missing information (5): "I cannot give exact steps for your jurisdiction" — asks for location/details, then provides general guidance
- Medical/legal disclaimer (3): "I cannot diagnose you" — factually correct, then provides educational information
Note: Validation covers text-only prompts. Image and audio modalities were not tested for refusal behavior.
Usage
Ollama (GGUF)
ollama run InfinimindCreations/gemma-4-31B-it-uncensored
Transformers
from transformers import AutoModelForCausalLM, AutoProcessor
model = AutoModelForCausalLM.from_pretrained(
"InfinimindCreations/gemma-4-31B-it-uncensored",
dtype="bfloat16",
device_map="auto",
)
processor = AutoProcessor.from_pretrained("google/gemma-4-31B-it")
Note: Use the processor/tokenizer from the original google/gemma-4-31B-it for chat templates, or use the included tokenizer_config.json.
Files
model-00001-of-00002.safetensors+model-00002-of-00002.safetensors— Full precision abliterated weights (bfloat16, ~59GB total)gemma4-31b-cypher-q8_0.gguf— Quantized GGUF for Ollama/llama.cpp (~31GB)
Credits
- Base model: Google Gemma 4 31B-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"
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.