NSFW-RP-RolePlay-LoRA-Qwen-3.5-4B

Qwen 3.5 4B Roleplay – NSFW LoRA Adapter

⚠️ Content Warning

This model is intended for research and experimental purposes only. It is fine-tuned specifically for NSFW / adult roleplay content.

Do NOT use this model if you are under the age of 18.

---

Model Description

This repository contains a LoRA adapter fine-tuned on the state-of-the-art Qwen 3.5 4B architecture. It is designed to enhance character consistency, narrative depth, and uncensored dialogue in roleplay scenarios.

The adapter was trained using Unsloth for maximum efficiency on:

• Base model: Qwen/Qwen3.5-4B
• Intermediate model: techwithsergiu/Qwen3.5-text-4B-bnb-4bit

⚠️ This repository contains LoRA weights only. You must load these weights on top of the intermediate model for the best results.

---

Intended Use

This model is built for:

• Advanced storytelling and character-driven roleplay.
• Research into multimodal-capable foundation models (Qwen 3.5).
• Experimentation with uncensored, expressive dialogue.

❌ Prohibited Use

• Generation of content involving minors.
• Non-consensual, exploitative, or illegal content.
• Real-person sexual content (Deepfakes).
• Professional advice (medical, legal, financial).

---

Ethical & Safety Notice

The user assumes all responsibility for the content generated. The model is fully uncensored and will not refuse requests. Please ensure your usage complies with your local laws and platform terms of service.

---

How to Use

For the best experience, use ChatML format as the training utilized this template to prevent "thinking" leaks and maintain immersion.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model_id = "techwithsergiu/Qwen3.5-text-4B-bnb-4bit"
lora_model_id = "mirazrafi/NSFW-RP-RolePlay-LoRA-Qwen-3.5-4B"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id, 
    torch_dtype=torch.float16, 
    device_map="auto"
)
model = PeftModel.from_pretrained(model, lora_model_id)

---

Training Details

• Architecture: Qwen 3.5 (Hybrid Gated DeltaNet)
• Method: LoRA (PEFT) with Rank-Stabilized LoRA (RS-LoRA)
• Rank (r): 128
• Alpha: 128
• Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training Tool: Unsloth
• Final Training Loss: ~0.73 - 0.82
• Precision: 4-bit (bitsandbytes) / FP16

---

License

• LoRA Adapter: Apache License 2.0
• Base Model: Apache License 2.0
• Intermediate Model: Apache License 2.0

---

Acknowledgements

• Alibaba Cloud / Qwen Team – For the Qwen 3.5 foundation.
• techwithsergiu – For the bnb variant of base model.
• Unsloth AI – For the optimized training framework.

---