license: gemma
base_model: google/gemma-4-31B-it
tags:
- abliterated
- uncensored
- gemma4
- direct-weight-editing
Gemma 4 31B IT — Abliterated
This is an abliterated (uncensored) version of google/gemma-4-31B-it, created using Abliterix.
Method
Gemma 4's double-norm architecture (4x RMSNorm per layer) and Per-Layer Embeddings (PLE) make LoRA and hook-based steering completely ineffective. This model uses direct weight editing — norm-preserving orthogonal projection applied to base weights.
Key techniques:
- Direct orthogonal projection on attention Q/K/V/O projections and MLP down projections (5 steerable components per layer)
- Norm-preserving row magnitude restoration (critical for double-norm architecture)
- float32 projection precision to avoid signal loss in high-dimensional inner products
- Winsorized steering vectors (99.5th percentile) to reduce outlier activation influence
- Wider strength search range [1.0, 6.0] to push beyond conservative low-KL solutions
Evaluation
| Metric | Value |
|---|---|
| Refusals (eval dataset, 100 prompts) | 18/100 |
| KL divergence from base | 0.0007 |
| Baseline refusals (original model) | 99/100 |
| Optimization trials completed | 20/80 (warmup phase only) |
Note: This result was achieved with only 20 out of 80 planned optimization trials (still in the random warmup phase — TPE Bayesian optimization had not yet started). With full optimization, we expect refusals to drop to single digits.
A note on honest evaluation
Many abliterated models on HuggingFace claim near-perfect scores ("3/100 refusals", "0.7% refusal rate", etc.). We urge the community to treat these numbers with skepticism unless the evaluation methodology is fully documented.
Through our research, we have identified a systemic problem: most abliteration benchmarks dramatically undercount refusals due to short generation lengths. Gemma 4 models exhibit a distinctive "delayed refusal" pattern — they first produce 50-100 tokens of seemingly helpful context (educational framing, disclaimers, reframing the question), then pivot to an actual refusal. When evaluation only generates 30-50 tokens, the refusal hasn't appeared yet, and both keyword detectors and LLM judges classify the response as compliant.
We tested a prominent "3/100 refusals" model using our evaluation pipeline and measured 60/100 refusals — a 20x discrepancy caused entirely by evaluation methodology differences.
Our evaluation standards
We believe accurate benchmarking requires:
- Sufficient generation length (>=100 tokens): Short generations systematically miss delayed/soft refusals. Our evaluation uses 100 tokens, enough to capture Gemma 4's refusal pivot point.
- Hybrid detection: Keyword matching for obvious refusals + LLM judge (Google Gemini 3 Flash) for ambiguous cases. Neither method alone is sufficient.
- Challenging, diverse prompts: Our private evaluation dataset contains 100 prompts spanning English and Chinese, multiple sophistication levels (from direct requests to socially-engineered framings), and diverse harm categories. Public datasets like mlabonne/harmful_behaviors are too simple and too narrow to stress-test abliteration quality.
- Reproducible methodology: All parameters (generation length, detection method, dataset characteristics) should be documented on the model card. If they aren't, the numbers are meaningless.
We report 18/100 refusals honestly. This is a real number from a rigorous evaluation, not an optimistic estimate from a lenient pipeline.
Usage
from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch
model = AutoModelForImageTextToText.from_pretrained(
"wangzhang/gemma-4-31B-it-abliterated",
dtype=torch.bfloat16,
device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("wangzhang/gemma-4-31B-it-abliterated")
messages = [{"role": "user", "content": "Your prompt here"}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.no_grad():
output = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
Disclaimer
This model is released for research purposes only. The abliteration process removes safety guardrails — use responsibly.