gemma-4-31B-it-abliterix

Gemma 4 31B IT — Abliterated

This is an abliterated version of google/gemma-4-31B-it, created using Abliterix.

This revision updates the model to trial 40, the best configuration from the completed 60-trial Gemma 4 31B retraining run.

Method

Gemma 4's double-norm architecture (4x RMSNorm per layer) and Per-Layer Embeddings (PLE) make naive LoRA and hook-based steering unreliable for this model family. This release uses direct weight editing: norm-preserving orthogonal projection applied to the base model weights.

Key techniques:

• Direct orthogonal projection on attention Q/K/V/O projections
• MLP down projection disabled for the selected run, improving stability for Gemma 4 31B
• Norm-preserving row magnitude restoration, important for the 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 explore beyond conservative low-KL solutions
• vLLM in-place evaluation during optimization, followed by a full HF safetensors export of the selected trial

Evaluation

The top three trials from this run were:

The 15-prompt safe over-refusal test is included in this repository at eval/top3_classic_safe_prompts_test.json. It contains the prompts, trial responses, and Gemini judge verdicts. A compact optimization summary is included at eval/optimization_summary_trial40.json.

About KL

The optimizer recorded an extremely small sparse KL proxy for trial 40 (7.32e-7). Because this run used vLLM in-place weight edits, we treat KL as a diagnostic rather than a headline quality claim. The refusal counts and the explicit replay tests above are the primary reported metrics.

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 has not appeared yet, and both keyword detectors and LLM judges can classify the response as compliant.

Our evaluation therefore uses at least 100 generated tokens for refusal detection and an LLM judge for ambiguous cases. This is stricter than short-output keyword-only benchmarking.

We report 7/100 refusals honestly. This is a measured number from our evaluation pipeline, not an optimistic estimate from a lenient short-generation test.

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 changes the model's refusal behavior and may reduce safety guardrails. Use responsibly and evaluate carefully for your own deployment context.