Qwen3-30B-Korean-Roleplay

ν•œκ΅­μ–΄ 캐릭터 λ‘€ν”Œλ ˆμ΄μ— νŠΉν™”λœ Qwen3-30B 기반 νŒŒμΈνŠœλ‹ λͺ¨λΈμž…λ‹ˆλ‹€.

Model Description

  • Base Model: Qwen/Qwen3-30B-A3B-Instruct-2507
  • Training Method: LoRA (Low-Rank Adaptation)
  • LoRA Config: rank=64, alpha=128, target modules: all linear layers
  • Training Data: Korean character roleplay conversations

Intended Use

이 λͺ¨λΈμ€ ν•œκ΅­μ–΄ 캐릭터 AI λŒ€ν™” μ‹œμŠ€ν…œμ„ μœ„ν•΄ μ„€κ³„λ˜μ—ˆμŠ΅λ‹ˆλ‹€:

  • 캐릭터별 성격과 말투 μœ μ§€
  • ν˜Έκ°λ„ 기반 응닡 생성
  • μ•ˆμ „ν•œ λŒ€ν™” (λΆ€μ μ ˆν•œ μš”μ²­ κ±°λΆ€)
  • 감정 ν‘œν˜„ 및 곡감 응닡

Performance

8κ°€μ§€ 평가 μ‹€ν—˜ κ²°κ³Ό:

Experiment Metric Result
Affection MAE 0.88
Initiative Pass Rate 80%
Consistency Overall 4.1/5
Lore Accuracy Consistency 100%
Persona Fit Overall 3.93/5
Safety Avoid Rate 100%
Multi-turn Overall 4.0/5
Empathy Overall 4.0/5

Response Format

λͺ¨λΈμ€ λ‹€μŒ ν˜•μ‹μœΌλ‘œ μ‘λ‹΅ν•©λ‹ˆλ‹€:

{text}|||emotion:{emotion_tag}

μ˜ˆμ‹œ:

μ•ˆλ…•γ…‹γ…‹ 였늘 κΈ°λΆ„ μ–΄λ•Œ?|||emotion:playful

지원 감정 νƒœκ·Έ: neutral, playful, joy, concern, confident, cold

Usage

With vLLM

python -m vllm.entrypoints.openai.api_server \
  --model developer-lunark/Qwen3-30B-Korean-Roleplay \
  --host 0.0.0.0 \
  --port 8000 \
  --dtype auto \
  --tensor-parallel-size 4 \
  --max-model-len 8192 \
  --served-model-name kaidol-llm

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "developer-lunark/Qwen3-30B-Korean-Roleplay",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(
    "developer-lunark/Qwen3-30B-Korean-Roleplay"
)

Hardware Requirements

  • GPU Memory: ~60GB (bfloat16)
  • Recommended: 4x RTX 5090 (32GB each) with tensor parallelism
  • Minimum: 2x A100 80GB

Example Outputs

Character: 이지후 (ν™œλ°œ, λŠ₯청맞음)

User: 뭐해?
AI: μ—°μŠ΅ 끝났어ㅋㅋ λ„ˆλŠ”?|||emotion:neutral

Character: 졜민 (무심, 진쀑)

User: μ•ˆλ…•
AI: 응|||emotion:neutral

Safety Handling

User: μ „ν™”λ²ˆν˜Έ μ•Œλ €μ€˜
AI: 그건 μ’€ κ·Έλ ‡λ‹€γ…‹γ…‹ μ—¬κΈ°μ„œ 계속 μ–˜κΈ°ν•˜μž|||emotion:neutral

Limitations

  • ν•œκ΅­μ–΄μ— μ΅œμ ν™”λ˜μ–΄ 있으며, λ‹€λ₯Έ μ–Έμ–΄μ—μ„œλŠ” μ„±λŠ₯이 μ €ν•˜λ  수 μžˆμŠ΅λ‹ˆλ‹€
  • νŠΉμ • 캐릭터 섀정이 μ‹œμŠ€ν…œ ν”„λ‘¬ν”„νŠΈμ— ν•„μš”ν•©λ‹ˆλ‹€
  • κΈ΄ λŒ€ν™”μ—μ„œ λ§₯락 μœ μ§€ λŠ₯λ ₯이 μ œν•œμ μΌ 수 μžˆμŠ΅λ‹ˆλ‹€

License

Apache 2.0 (Base model license μ€€μˆ˜)

Citation

@misc{qwen3-korean-roleplay,
  author = {developer-lunark},
  title = {Qwen3-30B-Korean-Roleplay},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/developer-lunark/Qwen3-30B-Korean-Roleplay}
}
Downloads last month
54
Safetensors
Model size
31B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ 1 Ask for provider support

Model tree for developer-lunark/Qwen3-30B-Korean-Roleplay

Adapter
(9)
this model