madokalif/dtol-claim-classifier-qwen3b

DTOL Claim Classifier - Qwen2.5-3B LoRA

건설장비 AS 클레임 원인 분류 모델

Model Details

• Base Model: Qwen/Qwen2.5-3B
• Fine-tuning: LoRA (r=32, alpha=64)
• Task: 27-class classification (cause analysis)
• Accuracy: 43.00%
• Top-5 Accuracy: 86.10%

Training Details

• Epochs: 10
• Batch Size: 4 (with gradient accumulation 4)
• Learning Rate: 1e-4 (cosine scheduler)
• Data: 4,000 training samples (imbalanced)

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel

# Load base model
base_model = AutoModelForSequenceClassification.from_pretrained(
    "Qwen/Qwen2.5-3B",
    num_labels=27,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "YOUR_USERNAME/dtol-claim-classifier-qwen3b")

# Tokenize and predict
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B")
inputs = tokenizer("결함 설명 텍스트", return_tensors="pt")
outputs = model(**inputs)

Labels (27 classes)

• Poor material
• Improper manufacturing
• Short
• Improper assembly
• Poor Design
• ... and 22 more

Note

이 모델은 ML+RAG 앙상블 (64.8%)보다 낮은 성능을 보입니다. 실제 서비스에는 ML+RAG 앙상블 파이프라인을 권장합니다.

License

MIT