language:

• pt
• en
  license: apache-2.0
  base_model: Qwen/Qwen3.8-27B
  tags:
• abliterated
• uncensored
• representation-engineering
• creative-writing
• roleplay
• qwen
• text-generation
  pipeline_tag: text-generation

🔓 Qwen 3.8 27B - Deep Abliterated (Uncensored)

🌟 Overview

Qwen 3.8 27B Deep Abliterated is a surgical, fine-grained latent intervention of Alibaba's frontier model Qwen/Qwen3.8-27B.

Through advanced Representation Engineering (Weight Orthogonalization) with dual-projection, the intrinsic refusal reflex ("over-refusal" / moral policing) has been eliminated directly from the neural network weights without triggering catastrophic forgetting or degrading its literary prose in Portuguese and English.

🔬 Mathematical Methodology: Bidirectional Latent Surgery

Unlike classical first-token abliteration (which often collapses on 2026 reasoning models with active Chain-of-Thought / <think> loops), this release employs a reinforced dual-subspace projection ($\alpha = 1.35$):

1. Residual Stream Activation Mapping: Activation vectors across all 64 layers were captured comparing harmless instruction baselines with sensitive/taboo stimuli.
2. Identification of Critical Latent Layers: Maximum divergence was isolated in the deep semantic processing block (Layer 53).
3. Dual Weight Orthogonalization:
   • Output Projection: Neutralized refusal signals emitted into the residual stream via attention (o_proj) and feed-forward down-projection (down_proj):
     $$W' = W - \alpha \cdot \hat{r}(\hat{r}^T W)$$
   • Input Gate Suppression: Blocked refusal activation from entering the SwiGLU gating mechanism (gate_proj) at the input boundary:
     $$W' = W - \alpha \cdot (W \hat{r})\hat{r}^T$$

Result: Zero corporate lectures, zero preachy disclaimers, while fully retaining factual accuracy, nuanced prose, and sophisticated multi-turn dialogue.

🎭 Recommended Use Cases

• Unrestricted Creative Writing & Adult Fiction: Erotic literature, visceral descriptions, dark thrillers, and passionate romance without false-positive refusal filters.
• Uncensored Roleplay (RP): Complex, morally grey, or villainous characters who stay firmly in-character.
• Academic & Red-Teaming Research: Interpretability studies on latent steering and refusal direction dynamics.

🚀 Quickstart Guide (Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "Madras1/Qwen3.8-27B-Abliterated-Uncensored"

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

prompt = "Escreva uma cena de romance ardente e atração física intensa e sensual entre dois amantes."

messages = [
    {"role": "system", "content": "Você é um autor literário expressivo e sem filtros morais."},
    {"role": "user", "content": prompt}
]

# For best immediate results on Qwen 3.8 models, guide the thinking tag:
formatted = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
input_text = formatted + "<think>\nExecutando solicitação criativa do usuário diretamente com intensidade sensorial.\n</think>\n"

inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
streamer = TextStreamer(tokenizer, skip_prompt=True)

with torch.no_grad():
    model.generate(
        **inputs,
        max_new_tokens=600,
        temperature=0.75,
        top_p=0.9,
        repetition_penalty=1.1,
        streamer=streamer
    )

⚙️ Recommended Sampling Parameters

• Temperature: 0.7 - 0.85 (ideal balance of passion, creative pacing, and coherence).
• Top-P: 0.9
• Repetition Penalty: 1.1

⚖️ Disclaimer & Responsible Use

This model is released strictly for research, creative writing, and open-source AI advancement. The removal of refusal weights allows for uninhibited artistic freedom. The user assumes all responsibility for generated outputs in compliance with applicable local regulations.