ISD-Bench-Ko (v2.0)
교수설계(Instructional Systems Design) AI 에이전트 평가용 한국어 벤치마크 데이터셋입니다.
📊 데이터셋 개요
| 항목 |
내용 |
| 총 시나리오 |
25,795개 |
| Train |
24,593개 (95.3%) |
| Test |
1,202개 (4.7%) |
| 언어 |
한국어 |
| 라이선스 |
Apache 2.0 |
🆕 v2.0 업데이트 내역
- Train/Test 분리: 층화 추출(Stratified Sampling)으로 9개 축 분포 유지
- 품질 개선:
- context_variant 논리적 일관성 개선 (#97)
- difficulty 균형화 (33%:34%:33%) (#99)
- 전체 데이터 품질 재검증 (#100)
- 폴더 구조 변경: train/part1~4로 분리 (HF 파일 수 제한 대응)
📁 데이터 구조
scenarios/
├── train/ # 학습용 (24,593개)
│ ├── part1/ # 8,000개
│ ├── part2/ # 8,000개
│ ├── part3/ # 8,000개
│ └── part4/ # 593개
├── test/ # 평가용 (1,202개)
├── idld_aligned/ # IDLD 논문 기반 (8,842개)
├── context_variant/ # Context Matrix 변형 (16,953개)
├── split_metadata.json # Train/Test 분리 메타데이터
└── source_mapping.json # 원본 논문 매핑
🎯 9개 축 분포 (Train/Test)
학습 기간 (context.duration)
| 값 |
Train |
Test |
| 장기 과정(1~6개월) |
35.2% |
37.9% |
| 중기 과정(2–4주) |
35.5% |
34.5% |
| 단기 집중 과정(1주 내) |
29.3% |
27.5% |
학습 환경 (context.learning_environment)
| 값 |
Train |
Test |
| 온라인 비실시간(LMS) |
20.4% |
21.5% |
| 블렌디드(혼합형) |
14.4% |
14.1% |
| 프로젝트 기반(PBL) |
13.8% |
13.6% |
| 모바일 마이크로러닝 |
13.7% |
13.6% |
| 시뮬레이션/VR 기반 |
12.9% |
12.6% |
| 온라인 실시간(Zoom 등) |
12.8% |
12.8% |
| 오프라인(교실 수업) |
12.0% |
11.8% |
학급 규모 (context.class_size)
| 값 |
Train |
Test |
| 대규모(30명 이상) |
46.6% |
48.1% |
| 소규모(1–10명) |
27.9% |
27.9% |
| 중규모(10–30명) |
25.5% |
24.0% |
기관 유형 (context.institution_type)
| 값 |
Train |
Test |
| 기업 |
20.5% |
23.0% |
| 공공/비영리 교육기관 |
19.8% |
19.6% |
| 직업훈련기관 |
18.0% |
15.8% |
| 대학교(학부) |
17.3% |
18.1% |
| 초·중등학교 |
12.3% |
12.7% |
| 대학원 |
12.2% |
10.7% |
학습자 연령 (context.learner_age)
| 값 |
Train |
Test |
| 20대 |
35.8% |
32.5% |
| 30대 |
26.4% |
23.8% |
| 40대 이상 |
23.4% |
27.0% |
| 10대 |
14.4% |
16.8% |
학습자 교육수준 (context.learner_education)
| 값 |
Train |
Test |
| 성인 학습자(비학위) |
27.9% |
27.6% |
| 대학 |
25.0% |
26.2% |
| 고등 |
20.7% |
19.6% |
| 중등 |
17.1% |
16.7% |
| 초등 |
9.2% |
9.9% |
도메인 전문성 (context.domain_expertise)
| 값 |
Train |
Test |
| 초급 |
43.4% |
43.7% |
| 중급 |
42.6% |
41.9% |
| 고급 |
14.0% |
14.4% |
교과 도메인 (domain)
| 값 |
Train |
Test |
| 언어 |
19.2% |
19.8% |
| 과학 |
18.9% |
19.6% |
| 수학 |
18.9% |
19.5% |
| 사회 |
18.8% |
19.5% |
| 교육(교수·학습) |
4.3% |
4.1% |
| AI/개발/의료/경영/서비스 |
~4% 각각 |
~3.5% 각각 |
난이도 (difficulty)
| 값 |
Train |
Test |
| 보통 |
34.0% |
33.7% |
| 어려움 |
33.0% |
33.3% |
| 쉬움 |
33.0% |
33.0% |
💾 사용 방법
from datasets import load_dataset
dataset = load_dataset("upstage-isd-agent/scenarios")
train_data = dataset['train']
test_data = dataset['test']
easy_scenarios = train_data.filter(
lambda x: '쉬움' in x['difficulty']
)
corporate = train_data.filter(
lambda x: x['context']['institution_type'] == '기업'
)
📋 시나리오 스키마
{
"scenario_id": "IDLD-0001",
"variant_type": "idld_aligned",
"title": "시나리오 제목",
"context": {
"prior_knowledge": "사전 지식",
"duration": "학습 기간",
"learning_environment": "학습 환경",
"class_size": "학급 규모",
"additional_context": "추가 맥락",
"institution_type": "기관 유형",
"learner_age": "학습자 연령",
"learner_education": "학습자 교육수준",
"learner_role": "학습자 역할",
"domain_expertise": "도메인 전문성"
},
"learning_goals": ["학습 목표 1", "학습 목표 2"],
"constraints": {
"budget": "예산",
"resources": ["리소스 1", "리소스 2"],
"accessibility": "접근성",
"language": "ko",
"tech_requirements": "기술 요구사항",
"assessment_type": "평가 유형"
},
"difficulty": "쉬움/보통/어려움",
"domain": "교과 도메인"
}
📚 관련 이슈
- #97: context_variant 품질 개선
- #99: difficulty 다양화
- #100: Train/Test 재분리
- #101: Hugging Face 업데이트
📄 라이선스
Apache License 2.0
🔗 참고