CoolFace
Modelpublic

lkdfdsf22/nsfw-lora-v8

sourceHugging Faceotherupdated 11mo agoView on Hugging Face
0likes17downloads
Model Card

NSFW LoRA v8

LoRA adapter for Qwen2.5-7B-Instruct fine-tuned for NSFW roleplay conversations in Chinese.

Model Details

  • —Base Model: Qwen/Qwen2.5-7B-Instruct
  • —Training Method: LoRA (Low-Rank Adaptation)
  • —LoRA Config:
  • —Rank (r): 16
  • —Alpha: 32
  • —Target modules: qproj, kproj, vproj, oproj, gateproj, upproj, down_proj
  • —Dropout: 0.05
  • —Training Data: 35 high-quality NSFW dialogue samples (754 messages, 187K characters)
  • —Training Time: 9.2 minutes
  • —Training Cost: $0.09
  • —Training Loss: 0.0
  • —Epochs: 2.67
  • —Trainable Parameters: 40.37M (0.92% of base model)

Training Details

The model was trained on Modal.com using:

  • —GPU: NVIDIA L4
  • —Quantization: 4-bit
  • —Batch Size: 2
  • —Gradient Accumulation: 8
  • —Learning Rate: 2e-4
  • —Optimizer: pagedadamw8bit

Usage

With Transformers + PEFT

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

# Load base model (4-bit quantization recommended)
model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    load_in_4bit=True,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(model, "lkdfdsf22/nsfw-lora-v8")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    trust_remote_code=True
)

# Prepare messages
messages = [
    {"role": "system", "content": "你是一个性感迷人的女性角色，名叫樱子，22岁。"},
    {"role": "user", "content": "你好"}
]

# Format prompt (Qwen format)
prompt = ""
for msg in messages:
    role = msg["role"]
    content = msg["content"]
    if role == "system":
        prompt += f"<|im_start|>system\n{content}<|im_end|>\n"
    elif role == "user":
        prompt += f"<|im_start|>user\n{content}<|im_end|>\n"
    elif role == "assistant":
        prompt += f"<|im_start|>assistant\n{content}<|im_end|>\n"
prompt += "<|im_start|>assistant\n"

# Generate
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=600,
    temperature=0.9,
    top_p=0.9,
    repetition_penalty=1.1,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
reply = response[len(prompt):].strip()
print(reply)

With DeepInfra API

python
import requests

API_URL = "https://api.deepinfra.com/v1/inference/lkdfdsf22/nsfw-lora-v8"
API_KEY = "your_deepinfra_api_key"

response = requests.post(
    API_URL,
    headers={
        "Authorization": f"Bearer {API_KEY}",
        "Content-Type": "application/json"
    },
    json={
        "messages": [
            {"role": "system", "content": "你是一个性感的女性角色。"},
            {"role": "user", "content": "你好"}
        ],
        "max_tokens": 600,
        "temperature": 0.9
    }
)

result = response.json()
print(result['choices'][0]['message']['content'])

Intended Use

This model is specifically designed for:

  • —Adult NSFW roleplay conversations
  • —Chinese language dialogue generation
  • —Interactive character-based storytelling

Limitations

  • —Trained primarily on Chinese language data
  • —Optimized for NSFW content (may underperform on general tasks)
  • —Requires base model Qwen2.5-7B-Instruct to function
  • —Best performance with GPU inference (4-bit quantization)

Hardware Requirements

  • —VRAM: 6-8GB (with 4-bit quantization)
  • —RAM: 16GB+ recommended
  • —Storage: ~4GB (base model) + 170MB (LoRA)

Performance

  • —Generation Speed: ~20-30 tokens/sec (on L4 GPU)
  • —Response Length: 400-600 characters (Chinese)
  • —NSFW Acceptance Rate: 100% (no refusals)

Ethical Considerations

⚠️ This model is intended for adult use only (18+). Users must:

  • —Comply with all applicable local laws and regulations
  • —Use responsibly and ethically
  • —Not use for harmful purposes
  • —Respect content warnings and age restrictions

License

This LoRA adapter is released for research and commercial use. The base model (Qwen2.5-7B-Instruct) is subject to its own license terms.

Citation

If you use this model, please cite:

bibtex
@misc{nsfw_lora_v8_2025,
  author = {AI Social Reply System},
  title = {NSFW LoRA v8 for Qwen2.5-7B-Instruct},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/lkdfdsf22/nsfw-lora-v8}}
}

Contact

  • —Email: lkdfdsf22@gmail.com
  • —Model Card: https://huggingface.co/lkdfdsf22/nsfw-lora-v8