Smoothie-Qwen3-30B-General-Roleplay-LoRA

A LoRA adapter fine-tuned on dnotitia/Smoothie-Qwen3-30B-A3B for general Korean roleplay and conversational AI tasks.

Model Description

This model is a LoRA (Low-Rank Adaptation) adapter designed to enhance roleplay and character-based conversation capabilities. It builds upon the Smoothie-Qwen3-30B-A3B base model, which combines the strengths of Qwen3-30B-A3B with Korean language optimizations.

Key Features

• Korean Language Optimized: Enhanced Korean language understanding and generation
• Roleplay Focused: Fine-tuned for character roleplay and immersive conversations
• Efficient Adaptation: Uses LoRA for memory-efficient fine-tuning
• Compatible with vLLM: Can be served with vLLM's LoRA adapter support

Usage

With Transformers + PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "dnotitia/Smoothie-Qwen3-30B-A3B",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("dnotitia/Smoothie-Qwen3-30B-A3B")

# Load LoRA adapter
model = PeftModel.from_pretrained(
    base_model,
    "developer-lunark/Smoothie-Qwen3-30B-General-Roleplay-LoRA"
)

# Generate
messages = [{"role": "user", "content": "안녕하세요!"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With vLLM (LoRA Adapter)

python -m vllm.entrypoints.openai.api_server \
  --model dnotitia/Smoothie-Qwen3-30B-A3B \
  --enable-lora \
  --lora-modules "roleplay-lora=developer-lunark/Smoothie-Qwen3-30B-General-Roleplay-LoRA" \
  --max-lora-rank 32 \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.85

Training Details

Training Configuration

Parameter	Value
Base Model	dnotitia/Smoothie-Qwen3-30B-A3B
Training Method	SFT (Supervised Fine-Tuning)
Precision	bfloat16
Learning Rate	1e-4
LR Scheduler	Cosine
Warmup Ratio	0.03
Weight Decay	0.01
Batch Size	4 (per device)
Gradient Accumulation	8
Effective Batch Size	32
Epochs	1
Total Steps	857

LoRA Configuration

Parameter	Value
LoRA Rank (r)	32
LoRA Alpha	64
LoRA Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj
Task Type	CAUSAL_LM

Training Metrics

Training progress and metrics can be viewed on Weights & Biases:

Framework Versions

• PEFT: 0.17.1
• TRL: 0.24.0
• Transformers: 4.57.1
• PyTorch: 2.8.0
• Datasets: 4.3.0
• Tokenizers: 0.22.1

Related Models

• developer-lunark/Qwen3-30B-Korean-Roleplay - Full merged model for direct inference

License

This model is released under the Apache 2.0 license, following the base model's license terms.

Citation

@misc{smoothie-qwen3-30b-roleplay-lora,
  title={Smoothie-Qwen3-30B-General-Roleplay-LoRA},
  author={developer-lunark},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/developer-lunark/Smoothie-Qwen3-30B-General-Roleplay-LoRA}
}

If you use TRL in your work, please cite:

@misc{vonwerra2022trl,
  title={{TRL: Transformer Reinforcement Learning}},
  author={Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
  year={2020},
  journal={GitHub repository},
  publisher={GitHub},
  howpublished={\url{https://github.com/huggingface/trl}}
}