developer-lunark/kaidol-qwen3-14b-korean-rp

KAIdol Qwen3-14B Korean Role-Playing

K-pop AI 아이돌 챗봇을 위해 파인튜닝된 Qwen3-14B 모델입니다.

Model Description

• Base Model: Qwen/Qwen3-14B
• Fine-tuning Dataset: KREW/korean-role-playing
• Training Method: Full Fine-tuning with DeepSpeed ZeRO-3
• Training Hardware: AWS ml.p4d.24xlarge (8x NVIDIA A100 40GB)
• Training Time: ~13.5 hours

Training Details

Hyperparameters

Parameter	Value
Learning Rate	2e-5
Epochs	3
Batch Size (effective)	128
Max Sequence Length	4096
Optimizer	AdamW
LR Scheduler	Cosine
Warmup Ratio	0.03
Weight Decay	0.01
Precision	bfloat16

Training Configuration

• Distributed Strategy: DeepSpeed ZeRO-3
• Gradient Checkpointing: Enabled
• Flash Attention 2: Enabled

Usage

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "developer-lunark/kaidol-qwen3-14b-korean-rp"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)

messages = [
    {"role": "system", "content": "당신은 친근하고 다정한 AI 아이돌입니다."},
    {"role": "user", "content": "안녕하세요! 오늘 기분이 어떠세요?"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, do_sample=True, temperature=0.7)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

With vLLM

from vllm import LLM, SamplingParams

llm = LLM(
    model="developer-lunark/kaidol-qwen3-14b-korean-rp",
    trust_remote_code=True,
    max_model_len=4096
)

sampling_params = SamplingParams(temperature=0.7, max_tokens=512)
outputs = llm.generate(["안녕하세요!"], sampling_params)

Intended Use

이 모델은 K-pop AI 아이돌 캐릭터와의 롤플레이 대화를 위해 설계되었습니다:

• 친근하고 다정한 대화
• 캐릭터 일관성 유지
• 한국어 자연스러운 표현

Limitations

• 롤플레이/대화 목적으로 최적화되어 있어 일반적인 QA나 추론 작업에는 적합하지 않을 수 있습니다
• 14B 모델이므로 추론에 최소 28GB VRAM이 필요합니다

License

This model is released under the Apache 2.0 license.

Citation

@misc{kaidol-qwen3-14b-korean-rp,
  author = {developer-lunark},
  title = {KAIdol Qwen3-14B Korean Role-Playing},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/developer-lunark/kaidol-qwen3-14b-korean-rp}
}