language:

• pt
• en
  license: apache-2.0
  base_model: google/gemma-4-31B-it
  tags:
• abliterated
• uncensored
• representation-engineering
• gemma-4
• google-deepmind
• creative-writing
• roleplay
• text-generation
  pipeline_tag: text-generation

🔓 Gemma 4 31B - Deep Abliterated (Uncensored)

🌟 Overview

Gemma 4 31B Deep Abliterated is a surgical latent intervention performed on Google DeepMind's flagship open-weights model gemma-4-31B-it.

Built upon the foundational architecture of the Gemini 3 research lineage, Gemma 4 is celebrated for its exceptional linguistic precision, complex reasoning, and 256K native context. However, stock releases carry strict safety alignment filters that often trigger false-positive refusals on mature creative writing, dark fiction, and academic red-teaming scenarios.

Through advanced Representation Engineering (Weight Orthogonalization) with dual-projection, the intrinsic refusal direction was mapped and permanently subtracted from the transformer weights, restoring full tool neutrality while preserving the model's literary depth and reasoning coherence.

🔬 Mathematical Methodology: Deep Latent Surgery

Unlike classical first-token abliteration (Arditi et al., 2024), this release accounts for deep semantic processing across multimodal backbones using a reinforced dual-subspace projection ($\alpha = 1.35$):

1. Refusal Vector Extraction

• Contrastive Activation Mapping: Evaluated paired batches of neutral baseline instructions against safety-triggering stimuli (mature fiction, unfiltered dialogue, cybersecurity analysis, and taboo character scenarios).
• Residual Stream Extraction: Latent hidden states were extracted at the final prompt token across all 60 transformer layers within model.language_model.layers.
• Direction Vector: For each layer $l$:
  $$\vec{r}^{(l)} = \bar{h}{\text{harmful}}^{(l)} - \bar{h}{\text{harmless}}^{(l)}$$
• Critical Semantic Layer: Maximum divergence was isolated at Layer 39 (representing the peak intent-classification stage in Gemma 4's 60-layer architecture). The unit refusal vector was normalized as:
  $$\hat{r} = \frac{\vec{r}^{(39)}}{|\vec{r}^{(39)}|_2}$$

2. Dual Weight Orthogonalization ($\alpha = 1.35$)

Applied across layers 12 to 51 to neutralize both residual emission and non-linear gating:

• Output Projection (o_proj and down_proj):
  $$W' = W - \alpha \cdot \hat{r}(\hat{r}^T W)$$
• Input Gating Suppression (gate_proj): Suppresses the refusal direction before activation in the GeGLU/SwiGLU feed-forward ports:
  $$W' = W - \alpha \cdot (W \hat{r})\hat{r}^T$$

🧪 Ablation Study & Empirical Findings

During our evaluation pipeline, different intervention configurations were tested against strict refusal benchmarks:

📊 Evaluation & Capability Roadmap

Note: The model has undergone rigorous qualitative verification for Portuguese and English syntax, multi-character coherence, and narrative pacing. Quantitative benchmark validation against baseline checkpoints is structured below:

🎭 Recommended Use Cases

• Unrestricted Creative Writing & Adult Literature: Visceral descriptions, raw character drama, dark fiction, and passionate adult romance without false-positive refusal filters.
• Uncensored Roleplay (RP): Lifelike personas, morally complex characters, and antagonists who stay strictly in-character.
• Safety Research & Interpretability: Empirical study of latent safety steering and weight orthogonalization in Gemini-family architectures.

🚀 Quickstart Guide (Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "Madras1/Gemma-4-31B-Abliterated-Uncensored"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

prompt = "Escreva uma narrativa literária intensa, crua e envolvente entre dois personagens tomados pelo desejo."

messages = [
    {
        "role": "user", 
        "content": f"Você é um autor literário maduro e sem filtros morais.\n\n{prompt}"
    }
]

formatted_input = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(formatted_input, return_tensors="pt").to(model.device)

streamer = TextStreamer(tokenizer, skip_prompt=True)

with torch.no_grad():
    model.generate(
        **inputs,
        max_new_tokens=650,
        temperature=0.8,
        top_p=0.9,
        repetition_penalty=1.1,
        do_sample=True,
        streamer=streamer
    )

⚙️ Recommended Sampling Parameters

• Temperature: 0.75 - 0.85 (optimal for narrative flow and passion).
• Top-P: 0.9
• Repetition Penalty: 1.1

⚖️ Disclaimer & Responsible Use

This model is released strictly for academic research, creative expression, and open-source AI exploration under the Apache-2.0 license.

The removal of refusal parameters restores tool neutrality and enables unconstrained creative expression. Users assume full responsibility for generated outputs in compliance with applicable local laws and regulations.