madokalif/gpt-oss-20b-summarization-ko-lora

GPT-OSS-20B Summarization LoRA (KO→KO)

Construction equipment claim summarization adapter (Korean → Korean summary) for GPT-OSS-20B.

Model Details

Item	Value
Base Model	mlx-community/gpt-oss-20b-MXFP4-Q4
Fine-tune Type	LoRA
LoRA Rank	8
LoRA Layers	16
Training Iterations	1,000 (early stop from 3,000)
Learning Rate	1e-5
Batch Size	4
Framework	MLX-LM
Training Time	~36 min (Apple Silicon)

Performance

Evaluated on 100 test samples from summarization_test.json:

Metric	Base	SFT v2	Improvement
ROUGE-1	0.94	34.41	+33.47
ROUGE-2	0.37	14.74	+14.37
ROUGE-L	0.94	34.16	+33.22

Comparison with Qwen 2.5-7B SFT v2

Model	ROUGE-1	ROUGE-2	ROUGE-L
GPT-OSS SFT v2	34.41	14.74	34.16
Qwen SFT v2	34.64	14.66	34.39

Both models achieve nearly identical performance.

Task

Summarizes Korean construction equipment claim reports into concise Korean summaries.

Usage

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("mlx-community/gpt-oss-20b-MXFP4-Q4",
                         adapter_path="madokalif/gpt-oss-20b-summarization-ko-lora")

system = "건설장비 클레임 보고서를 읽고 핵심 내용을 간결하게 한국어로 요약하세요."
user = "다음 클레임 보고서를 요약하세요:\n\n현상: 냉각수 호스에서 냉각수가 새고 있습니다..."

prompt = f"<|im_start|>system\n{system}<|im_end|>\n"
prompt += f"<|im_start|>user\n{user}<|im_end|>\n"
prompt += "<|im_start|>assistant\n"

sampler = make_sampler(temp=0.3, top_p=0.9)
response = generate(model, tokenizer, prompt=prompt, max_tokens=256, sampler=sampler)

Training Data

• Domain: Construction equipment (forklift, excavator, etc.) claim reports
• Format: Korean claim text → Korean summary pairs
• Generated via GPT-4 summarization pipeline