developer-lunark/kaidol-phase2a-test-1k

KAIDOL LLM Fine-tuning - PHASE2A-TEST-1K

Korean AI Idol Roleplay Language Model based on unsloth/Qwen3-30B-A3B-Instruct-2507

Model Description

이 모델은 K-pop 아이돌 스타일의 롤플레잉 및 공감 대화를 위해 fine-tuning된 LoRA adapter입니다.

• Base Model: unsloth/Qwen3-30B-A3B-Instruct-2507
• Training Phase: phase2a-test-1k
• Training Framework: Unsloth 2025.11.3
• LoRA Rank: 16
• LoRA Alpha: 16
• Training Samples: 1000

Training Configuration

{
  "model": "Qwen3-30B-A3B-Instruct-2507",
  "phase": "phase2a-test-1k",
  "dataset": "phase2-rp-base-1k",
  "num_samples": 1000,
  "lora_rank": 16,
  "lora_alpha": 16,
  "lora_dropout": 0,
  "learning_rate": 0.0002,
  "batch_size": 2,
  "gradient_accumulation_steps": 4,
  "effective_batch_size": 32,
  "max_steps": 100,
  "warmup_steps": 10,
  "max_seq_length": 2048,
  "optimizer": "adamw_8bit",
  "weight_decay": 0.01,
  "lr_scheduler_type": "linear",
  "precision": "bfloat16",
  "device_map": "auto",
  "gpus": "4x RTX 5090",
  "training_time": "40 minutes",
  "framework": "Unsloth 2025.11.3",
  "target_modules": [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj"
  ]
}

Evaluation Metrics

{
  "training_loss": {
    "initial": 2.3745,
    "final": 1.5027,
    "reduction_percent": 36.7
  },
  "training_metrics": {
    "total_steps": 100,
    "total_samples": 1000,
    "training_time_seconds": 2380.49,
    "training_time_minutes": 39.67,
    "samples_per_second": 0.336,
    "final_grad_norm": 0.1539,
    "final_learning_rate": 0.0
  },
  "loss_progression": {
    "step_5": 2.3745,
    "step_10": 1.531,
    "step_50": 1.632,
    "step_100": 1.5027
  },
  "wandb_run": "https://wandb.ai/developer_lunark-lunark-ai/kaidol-llm-finetuning/runs/brryct5m",
  "notes": "Baseline test with 1K samples. Stable convergence observed. Ready for hyperparameter optimization (LR 2e-4→1e-4, alpha 16→32, grad_accum 4→8)."
}

Usage

로드 방법 (Unsloth 사용)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="developer-lunark/kaidol-phase2a-test-1k",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

추론 예시

messages = [
    {"role": "user", "content": "오늘 기분이 좋지 않아..."},
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to("cuda")

outputs = model.generate(
    inputs,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.9,
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Dataset

• Phase 2: RP Base Dataset (54K samples)
  • Source: developer-lunark/kaidol-phase2-rp-base-v0.1
  • Korean: 53% / English: 47%

Training Hardware

• GPU: 4x NVIDIA RTX 5090 (32GB each)
• Training Time: ~40 minutes
• Framework: Unsloth + PyTorch 2.9.1 + CUDA 12.8

Limitations

• 이 모델은 롤플레잉 및 공감 대화에 특화되어 있습니다
• 일반적인 지식 질문이나 reasoning 작업에는 베이스 모델보다 성능이 낮을 수 있습니다
• 한국어와 영어 외의 언어는 제한적으로 지원됩니다

Ethical Considerations

• 이 모델은 연구 및 교육 목적으로 제작되었습니다
• 상업적 사용 시 라이선스를 확인하세요
• 생성된 콘텐츠의 품질과 적절성을 항상 검증하세요

Citation

@misc{kaidol-phase2a-test-1k,
  author = {Developer Lunark},
  title = {KAIDOL LLM Fine-tuning - PHASE2A-TEST-1K},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/developer-lunark/kaidol-phase2a-test-1k}}
}

Model Card Contact

• Developer: developer_lunark
• Repository: https://github.com/developer-lunark/kaidol-llm-finetuning
• W&B Project: https://wandb.ai/developer_lunark-lunark-ai/kaidol-llm-finetuning

---

Generated on 2025-11-18 09:24:35