madokalif/qwen2.5-7b-construction-summarization-ko-lora

Qwen2.5-7B Construction Equipment Summarization LoRA (Korean) ✅ Best

건설장비 유지보수 클레임 요약 LoRA 어댑터 (SFT v2 - 최종 추천 모델)

Model Description

• Base Model: Qwen/Qwen2.5-7B-Instruct
• Training Method: SFT v2 (Supervised Fine-Tuning)
• Domain: Construction equipment maintenance & repair claim summarization
• Task: Korean claim report → Korean summary
• Framework: Transformers + PEFT

Performance (summarization_test.json, 100 samples)

Model	ROUGE-1	ROUGE-2	ROUGE-L
Qwen Base	32.94	14.57	32.94
Qwen SFT v2 (this)	34.64	14.66	34.39
Qwen DPO v2	33.20	12.07	33.20

Comparison with GPT-OSS-20B

Model	ROUGE-1	ROUGE-2	ROUGE-L
Qwen SFT v2 (7B, this)	34.64	14.66	34.39
GPT-OSS SFT v2 (20B)	34.41	14.74	34.16

Nearly identical performance despite Qwen being 7B vs GPT-OSS 20B.

Key Findings

• Qwen Base already has strong Korean summarization ability (ROUGE-1 32.94)
• SFT provides modest improvement (+1.70 ROUGE-1)
• DPO hurts performance (-1.44 ROUGE-1 vs SFT)
• 100% of samples produce valid Korean summaries

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "madokalif/qwen2.5-7b-construction-summarization-ko-lora")

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

messages = [
    {"role": "system", "content": "건설장비 클레임 보고서를 읽고 핵심 내용을 간결하게 한국어로 요약하세요."},
    {"role": "user", "content": "다음 클레임 보고서를 요약하세요:\n\n현상: 냉각수 호스에서 냉각수가 새고 있습니다..."}
]

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=256, temperature=0.3, top_p=0.9, do_sample=True)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))