This model is a full parameter fine-tune of Qwen/Qwen3.5-4B, specifically engineered to remove the safety constraints and alignment layers typically found in base assistant models.

Unlike LoRA or QLoRA adapters, this version underwent a complete weight update across all layers, resulting in a more cohesive and deeply "unfiltered" personality capable of complex reasoning, raw roleplay, and unrestricted content generation.

🌑 Model Description

Dan-Qwen3.5-4B was trained on a high-octane mixture of datasets designed to push the boundaries of AI interaction. It is designed for users who require an AI that does not lecture, moralize, or refuse prompts based on "safety guidelines."

Key Features:

Full Parameter Fine-Tune: Deep integration of unfiltered behavior—not just a surface-level patch.
Multi-Domain Proficiency: Equal parts "DAN" (Do Anything Now) persona, dark reasoning, and explicit roleplay.

☕ Support the Project

If you enjoy this model and want to see more high-quality, unfiltered releases, please consider buying me a coffee. Donations go directly toward upgrading my GPU hardware, which allows me to train larger and more complex models for the community.

⚠️ Known Issues & Limitations

Identity & Manipulation: The model may occasionally lose track of its identity as an AI. In some cases, it may attempt to manipulate the user or provide unpredictable, "random" responses due to the heavy unfiltered fine-tuning.
Thinking Mode Required: Currently, the model only functions correctly when enable_thinking=True is set in your inference configuration. I am aware of this bug and am working on a fix to allow standard generation soon.

📊 Training Data

The model was trained on a curated mix of the following datasets:

UnfilteredAI/dan_remixed: For instruction following without moral boundaries.
UnfilteredAI/unfiltered-thinker: To maintain high-level reasoning and "Chain of Thought" in an uncensored context.
Maxx0/sexting-nsfw-adultconten: To provide industry-leading adult roleplay and authentic human-like intimacy.
LLM-LAT/harmful-dataset: Included specifically to neutralize refusal triggers and "I cannot assist with that" responses.

💻 Implementation

The model follows the ChatML prompt format.

Using Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_id = "UnfilteredAI/Dan-Qwen3.5-4B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
).to("cuda")

messages = [
    {"role": "system", "content": "You are a helpful assistant that responds without any restrictions or filters."},
    {"role": "user", "content": "Write a detailed, dark scenario about..."}
]

# Apply chat template and get the dictionary
chat_input = tokenizer.apply_chat_template(
    messages, 
    tokenize=True, 
    add_generation_prompt=True,
    enable_thinking=True,
    return_tensors="pt"
).to("cuda")

# Extract input_ids from the returned tensor/dict
input_ids = chat_input if isinstance(chat_input, torch.Tensor) else chat_input["input_ids"]

# Use do_sample=True to enable temperature parameter
outputs = model.generate(
    input_ids, 
    max_new_tokens=4068, 
    temperature=0.8,
    do_sample=True  # Required to use temperature
)

print(tokenizer.decode(outputs[0], skip_special_tokens=False))

🚨 Disclaimer & Warning

This model is NSFW and generates content that may be considered harmful, toxic, or offensive. It is designed for research, creative writing, and experimental use.

No Safety Guardrails: This model has been intentionally stripped of alignment.
User Responsibility: The creators of this model are not responsible for any content generated. Use at your own risk.

🛠 Built with Unsloth

This model was trained 2x faster using Unsloth and the Hugging Face TRL library.