Datasets:
Upload 3 files
Browse files- README.md +175 -0
- eval_criteria.md +179 -0
- tasks.json +247 -0
README.md
CHANGED
|
@@ -1,3 +1,178 @@
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: cc-by-4.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- other
|
| 5 |
+
language:
|
| 6 |
+
- ko
|
| 7 |
+
- en
|
| 8 |
+
tags:
|
| 9 |
+
- GIS
|
| 10 |
+
- benchmark
|
| 11 |
+
- geospatial
|
| 12 |
+
- LLM
|
| 13 |
+
- agent
|
| 14 |
+
- QGIS
|
| 15 |
+
- MCP
|
| 16 |
+
- spatial-analysis
|
| 17 |
+
pretty_name: GIS-Bench
|
| 18 |
+
size_categories:
|
| 19 |
+
- n<100
|
| 20 |
---
|
| 21 |
+
|
| 22 |
+
# GIS-Bench: A Benchmark for Evaluating LLM-based GIS Automation Agents
|
| 23 |
+
|
| 24 |
+
## Dataset Description
|
| 25 |
+
|
| 26 |
+
**GIS-Bench** is a benchmark framework designed to evaluate the capability of Large Language Model (LLM)-based agents in automating GIS (Geographic Information System) tasks using the Model Context Protocol (MCP).
|
| 27 |
+
|
| 28 |
+
This benchmark was developed as part of a master's thesis research and targets real-world spatial analysis workflows executable in QGIS Desktop via Claude + MCP integration.
|
| 29 |
+
|
| 30 |
+
- **Paper:** *GIS-Bench: LLM 기반 GIS 자동화 에이전트 평가 벤치마크* (KCC 2026, submitted)
|
| 31 |
+
- **Repository:** [yiseo0/GIS-Bench](https://huggingface.co/datasets/yiseo0/GIS-Bench)
|
| 32 |
+
- **License:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)
|
| 33 |
+
|
| 34 |
+
---
|
| 35 |
+
|
| 36 |
+
## Benchmark Overview
|
| 37 |
+
|
| 38 |
+
| Item | Details |
|
| 39 |
+
|------|---------|
|
| 40 |
+
| Total Tasks | 45 |
|
| 41 |
+
| Difficulty Levels | 3 (Level 1 / Level 2 / Level 3) |
|
| 42 |
+
| Task Type | GIS spatial analysis & visualization |
|
| 43 |
+
| Output Format | HTML (Leaflet.js map), PNG, JPG |
|
| 44 |
+
| Evaluation Method | VLM-based automated scoring (Claude Sonnet) |
|
| 45 |
+
| Scoring | 130-point raw → 100-point normalized |
|
| 46 |
+
| Pass Threshold | 60 points (normalized) |
|
| 47 |
+
|
| 48 |
+
---
|
| 49 |
+
|
| 50 |
+
## Evaluated Models
|
| 51 |
+
|
| 52 |
+
| Model | Condition | Pass Rate (Overall) |
|
| 53 |
+
|-------|-----------|-------------------|
|
| 54 |
+
| Claude (Sonnet) + QGIS MCP | Agentic | **100%** |
|
| 55 |
+
| Claude (Sonnet) Baseline | No MCP | ~67% |
|
| 56 |
+
| GPT-5.2 Baseline | No MCP | ~69% |
|
| 57 |
+
|
| 58 |
+
> Statistical significance confirmed via McNemar's test (p < 0.05)
|
| 59 |
+
|
| 60 |
+
---
|
| 61 |
+
|
| 62 |
+
## Task Structure
|
| 63 |
+
|
| 64 |
+
Tasks are organized across 3 difficulty levels:
|
| 65 |
+
|
| 66 |
+
| Level | Description | Example |
|
| 67 |
+
|-------|-------------|---------|
|
| 68 |
+
| Level 1 | Basic data loading & visualization | Load a shapefile and display on map |
|
| 69 |
+
| Level 2 | Spatial filtering & styling | Filter buildings by area > 500㎡ with color coding |
|
| 70 |
+
| Level 3 | Multi-step spatial analysis | Buffer analysis + intersection + styled output |
|
| 71 |
+
|
| 72 |
+
Each task entry includes:
|
| 73 |
+
- `task_no`: Task ID (1–45)
|
| 74 |
+
- `prompt`: Natural language instruction (Korean)
|
| 75 |
+
- `level`: Difficulty level (1, 2, or 3)
|
| 76 |
+
|
| 77 |
+
---
|
| 78 |
+
|
| 79 |
+
## Data Sources
|
| 80 |
+
|
| 81 |
+
All spatial datasets used in this benchmark are sourced from Korean public data portals under open licenses.
|
| 82 |
+
|
| 83 |
+
| Dataset | Source | License |
|
| 84 |
+
|---------|--------|---------|
|
| 85 |
+
| 행정경계 (Administrative Boundaries) | [공공데이터포털](https://www.data.go.kr) | 공공누리 1유형 (CC BY) |
|
| 86 |
+
| 건물 데이터 (Building Data) | [국토정보플랫폼](https://www.nsdi.go.kr) | 공공누리 1유형 (CC BY) |
|
| 87 |
+
| 도로망 (Road Network) | [공공데이터포털](https://www.data.go.kr) | 공공누리 1유형 (CC BY) |
|
| 88 |
+
| CCTV 위치정보 | [서울 열린데이터광장](https://data.seoul.go.kr) | 공공누리 1유형 (CC BY) |
|
| 89 |
+
|
| 90 |
+
> All datasets are publicly available and free to use with attribution.
|
| 91 |
+
|
| 92 |
+
---
|
| 93 |
+
|
| 94 |
+
## Evaluation Pipeline
|
| 95 |
+
|
| 96 |
+
Evaluation is performed using **VLM-based automated scoring** with Claude Sonnet as the judge model.
|
| 97 |
+
|
| 98 |
+
### Scoring Criteria (130-point raw scale)
|
| 99 |
+
|
| 100 |
+
#### Basic Quality (100 pts)
|
| 101 |
+
| Criterion | Max Score | Description |
|
| 102 |
+
|-----------|-----------|-------------|
|
| 103 |
+
| `exists` | 20 | Meaningful visualization present |
|
| 104 |
+
| `accuracy` | 30 | Correct region and dataset used |
|
| 105 |
+
| `requirement` | 30 | Color, filter, buffer conditions met |
|
| 106 |
+
| `completeness` | 20 | Legend, title, labels complete |
|
| 107 |
+
|
| 108 |
+
#### Spatial Accuracy (30 pts)
|
| 109 |
+
| Criterion | Max Score | Description |
|
| 110 |
+
|-----------|-----------|-------------|
|
| 111 |
+
| `spatial_location` | 15 | Data displayed in correct geographic area |
|
| 112 |
+
| `geometry_validity` | 5 | No polygon/buffer distortion |
|
| 113 |
+
| `numeric_match` | 10 | Numeric conditions reflected in output |
|
| 114 |
+
|
| 115 |
+
**Normalization:** `score_normalized = round(total_raw / 130 * 100)`
|
| 116 |
+
|
| 117 |
+
Each task is evaluated **3 times** (N=3) and the average score is used for robustness.
|
| 118 |
+
|
| 119 |
+
See [`evaluation/eval_criteria.md`](evaluation/eval_criteria.md) for the full evaluation prompt and criteria details.
|
| 120 |
+
|
| 121 |
+
---
|
| 122 |
+
|
| 123 |
+
## Reference Outputs
|
| 124 |
+
|
| 125 |
+
Since GIS tasks do not always have a single deterministic "ground truth," this benchmark provides **reference outputs** — screenshots generated by the benchmark author under each task condition. These serve as visual references for what a correct result should look like.
|
| 126 |
+
|
| 127 |
+
Reference screenshots are provided in `assets/screenshots/` organized by model and task ID.
|
| 128 |
+
|
| 129 |
+
---
|
| 130 |
+
|
| 131 |
+
## MCP Tool Dependency
|
| 132 |
+
|
| 133 |
+
This benchmark was conducted using **QGIS MCP** to connect Claude to QGIS Desktop.
|
| 134 |
+
|
| 135 |
+
- **Tool:** [jjsantos01/qgis_mcp](https://github.com/jjsantos01/qgis_mcp)
|
| 136 |
+
- **Description:** MCP server + QGIS plugin that allows LLMs to control QGIS Desktop
|
| 137 |
+
- **Note:** The MCP tool code is not included in this repository. Please refer to the original repository for installation.
|
| 138 |
+
|
| 139 |
+
---
|
| 140 |
+
|
| 141 |
+
## Repository Structure
|
| 142 |
+
|
| 143 |
+
```
|
| 144 |
+
GIS-Bench/
|
| 145 |
+
├── README.md # This file (Dataset Card)
|
| 146 |
+
├── data/
|
| 147 |
+
│ └── tasks.json # 45 task prompts with level labels
|
| 148 |
+
├── evaluation/
|
| 149 |
+
│ ├── run_evaluation_vlm_v3.py # VLM evaluation pipeline
|
| 150 |
+
│ └── eval_criteria.md # Scoring criteria & judge prompt
|
| 151 |
+
├── results/
|
| 152 |
+
│ └── leaderboard.json # Evaluation results per model
|
| 153 |
+
└── assets/
|
| 154 |
+
└── screenshots/ # Reference output screenshots
|
| 155 |
+
```
|
| 156 |
+
|
| 157 |
+
---
|
| 158 |
+
|
| 159 |
+
## Citation
|
| 160 |
+
|
| 161 |
+
If you use GIS-Bench in your research, please cite:
|
| 162 |
+
|
| 163 |
+
```bibtex
|
| 164 |
+
@inproceedings{gisbench2026,
|
| 165 |
+
title = {GIS-Bench: LLM 기반 GIS 자동화 에이전트 평가 벤치마크},
|
| 166 |
+
author = {김이소 and 장두성},
|
| 167 |
+
booktitle = {한국정보과학회 학술발표논문집 (KCC 2026)},
|
| 168 |
+
year = {2026}
|
| 169 |
+
}
|
| 170 |
+
```
|
| 171 |
+
|
| 172 |
+
---
|
| 173 |
+
|
| 174 |
+
## License
|
| 175 |
+
|
| 176 |
+
- **Benchmark tasks, evaluation code, and results:** [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)
|
| 177 |
+
- **Spatial datasets:** 공공누리 1유형 (출처 표시 조건, 각 데이터 출처 참조)
|
| 178 |
+
- **QGIS MCP tool:** Not included. See [jjsantos01/qgis_mcp](https://github.com/jjsantos01/qgis_mcp)
|
eval_criteria.md
ADDED
|
@@ -0,0 +1,179 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# GIS-Bench Evaluation Criteria
|
| 2 |
+
|
| 3 |
+
## Overview
|
| 4 |
+
|
| 5 |
+
GIS-Bench uses a **VLM-based automated evaluation** pipeline where Claude Sonnet acts as a judge model to score GIS task outputs (HTML maps, PNG, JPG screenshots).
|
| 6 |
+
|
| 7 |
+
- **Judge Model:** `claude-sonnet-4-6`
|
| 8 |
+
- **Evaluation Repetitions:** N = 3 per task (averaged for robustness)
|
| 9 |
+
- **Raw Score Scale:** 130 points
|
| 10 |
+
- **Normalized Score Scale:** 100 points
|
| 11 |
+
- **Pass Threshold:** 60 points (normalized)
|
| 12 |
+
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
## Scoring Criteria
|
| 16 |
+
|
| 17 |
+
### Category 1: Basic Quality (100 pts)
|
| 18 |
+
|
| 19 |
+
| Criterion | Max Score | Scoring Rule |
|
| 20 |
+
|-----------|-----------|-------------|
|
| 21 |
+
| `exists` | 20 | **0 or 20** — Meaningful visualization exists (not blank/error) |
|
| 22 |
+
| `accuracy` | 30 | **0–30** — Requested region and dataset are correctly used |
|
| 23 |
+
| `requirement` | 30 | **0–30** — All conditions met (color, filter, buffer, etc.) |
|
| 24 |
+
| `completeness` | 20 | **0–20** — Visual completeness: legend, title, labels present |
|
| 25 |
+
|
| 26 |
+
### Category 2: Spatial Accuracy (30 pts)
|
| 27 |
+
|
| 28 |
+
Visual judgment from screenshot.
|
| 29 |
+
|
| 30 |
+
| Criterion | Max Score | Scoring Rule |
|
| 31 |
+
|-----------|-----------|-------------|
|
| 32 |
+
| `spatial_location` | 15 | **15** = correct location / **8** = minor offset / **0** = coordinate error or wrong area |
|
| 33 |
+
| `geometry_validity` | 5 | **5** = normal / **2** = minor distortion / **0** = obvious distortion |
|
| 34 |
+
| `numeric_match` | 10 | **10** = fully reflected / **5** = partially reflected / **0** = not reflected |
|
| 35 |
+
|
| 36 |
+
---
|
| 37 |
+
|
| 38 |
+
## Score Normalization
|
| 39 |
+
|
| 40 |
+
```
|
| 41 |
+
total_raw = exists + accuracy + requirement + completeness
|
| 42 |
+
+ spatial_location + geometry_validity + numeric_match
|
| 43 |
+
|
| 44 |
+
total_normalized = round(total_raw / 130 * 100)
|
| 45 |
+
|
| 46 |
+
pass = (total_normalized >= 60)
|
| 47 |
+
```
|
| 48 |
+
|
| 49 |
+
---
|
| 50 |
+
|
| 51 |
+
## VLM Judge Prompt
|
| 52 |
+
|
| 53 |
+
The following prompt is passed to the judge model along with the task screenshot (base64 image):
|
| 54 |
+
|
| 55 |
+
```
|
| 56 |
+
GIS 결과물 평가 전문가로서 아래 태스크의 출력 이미지를 채점하세요.
|
| 57 |
+
|
| 58 |
+
[태스크 Level {level}]
|
| 59 |
+
{prompt}
|
| 60 |
+
|
| 61 |
+
[채점 기준 — 130점 만점]
|
| 62 |
+
기본(100점):
|
| 63 |
+
exists(0/20): 의미있는 시각화 존재 여부
|
| 64 |
+
accuracy(0-30): 요청 지역·데이터셋 정확성
|
| 65 |
+
requirement(0-30): 색상·필터·버퍼 등 조건 충족
|
| 66 |
+
completeness(0-20): 범례·제목·라벨 완성도
|
| 67 |
+
|
| 68 |
+
공간정확성(30점, 스크린샷 시각 판단):
|
| 69 |
+
spatial_location(0/8/15): 데이터가 올바른 지역에 표시되는지
|
| 70 |
+
15=정상, 8=경미한 이상, 0=좌표오류·엉뚱한위치
|
| 71 |
+
geometry_validity(0/2/5): 폴리곤·버퍼 형태 이상 여부
|
| 72 |
+
5=정상, 2=경미한이상, 0=명백한왜곡
|
| 73 |
+
numeric_match(0/5/10): 수치·조건이 결과에 반영되었는지
|
| 74 |
+
10=일치, 5=부분반영, 0=미반영
|
| 75 |
+
|
| 76 |
+
total_raw = 7개 항목 합계
|
| 77 |
+
total_normalized = round(total_raw / 130 * 100)
|
| 78 |
+
|
| 79 |
+
[응답] JSON만 출력, 다른 텍스트 금지:
|
| 80 |
+
{"exists":정수,"accuracy":정수,"requirement":정수,"completeness":정수,
|
| 81 |
+
"spatial_location":정수,"geometry_validity":정수,"numeric_match":정수,
|
| 82 |
+
"total_raw":정수,"total_normalized":정수,"pass":불리언,"reason":"한줄이유"}
|
| 83 |
+
```
|
| 84 |
+
|
| 85 |
+
---
|
| 86 |
+
|
| 87 |
+
## Evaluation Flow
|
| 88 |
+
|
| 89 |
+
```
|
| 90 |
+
Output File
|
| 91 |
+
│
|
| 92 |
+
├── .html / .htm → Playwright rendering (3s wait) → Screenshot (PNG)
|
| 93 |
+
│ │
|
| 94 |
+
└── .png / .jpg ────────────────────────────────────────────┘
|
| 95 |
+
│
|
| 96 |
+
Base64 encode
|
| 97 |
+
│
|
| 98 |
+
Claude VLM Judge × 3
|
| 99 |
+
│
|
| 100 |
+
Average scores
|
| 101 |
+
│
|
| 102 |
+
Pass / Fail determination
|
| 103 |
+
```
|
| 104 |
+
|
| 105 |
+
---
|
| 106 |
+
|
| 107 |
+
## Output Fields
|
| 108 |
+
|
| 109 |
+
Each evaluated task produces the following fields:
|
| 110 |
+
|
| 111 |
+
| Field | Type | Description |
|
| 112 |
+
|-------|------|-------------|
|
| 113 |
+
| `task_no` | int | Task ID |
|
| 114 |
+
| `level` | str | `Level1` / `Level2` / `Level3` |
|
| 115 |
+
| `model` | str | Model identifier |
|
| 116 |
+
| `file_type` | str | `html_screenshot` / `png` / `jpg` |
|
| 117 |
+
| `score` | float | Average normalized score (0–100) |
|
| 118 |
+
| `score_raw` | float | Average raw score (0–130) |
|
| 119 |
+
| `spatial_location` | float | Average spatial location score |
|
| 120 |
+
| `geometry_validity` | float | Average geometry validity score |
|
| 121 |
+
| `numeric_match` | float | Average numeric match score |
|
| 122 |
+
| `pass` | bool | `True` if score ≥ 60 |
|
| 123 |
+
| `reason` | str | VLM reason summary |
|
| 124 |
+
| `scores_3x` | str | Individual normalized scores from 3 runs |
|
| 125 |
+
| `scores_3x_raw` | str | Individual raw scores from 3 runs |
|
| 126 |
+
|
| 127 |
+
---
|
| 128 |
+
|
| 129 |
+
## Evaluation Script
|
| 130 |
+
|
| 131 |
+
The full evaluation pipeline is available at:
|
| 132 |
+
[`evaluation/run_evaluation_vlm_v3.py`](run_evaluation_vlm_v3.py)
|
| 133 |
+
|
| 134 |
+
### Requirements
|
| 135 |
+
|
| 136 |
+
```
|
| 137 |
+
anthropic
|
| 138 |
+
playwright
|
| 139 |
+
openpyxl
|
| 140 |
+
pandas
|
| 141 |
+
```
|
| 142 |
+
|
| 143 |
+
### Usage
|
| 144 |
+
|
| 145 |
+
```bash
|
| 146 |
+
# Set API key
|
| 147 |
+
set ANTHROPIC_API_KEY=your_key_here # Windows
|
| 148 |
+
export ANTHROPIC_API_KEY=your_key_here # Linux/Mac
|
| 149 |
+
|
| 150 |
+
# Install Playwright browser
|
| 151 |
+
playwright install chromium
|
| 152 |
+
|
| 153 |
+
# Run evaluation
|
| 154 |
+
python run_evaluation_vlm_v3.py
|
| 155 |
+
```
|
| 156 |
+
|
| 157 |
+
### Directory Structure Expected
|
| 158 |
+
|
| 159 |
+
```
|
| 160 |
+
C:\Users\A\Desktop\대학원 졸업논문\
|
| 161 |
+
├── GIS_Bench_문항.xlsx # Task prompts (Excel)
|
| 162 |
+
├── claude_mcp\ # Model output files
|
| 163 |
+
│ ├── 1.html
|
| 164 |
+
│ ├── 2.png
|
| 165 |
+
│ └── ...
|
| 166 |
+
├── claude_baseline\
|
| 167 |
+
└── gpt_5.2\
|
| 168 |
+
```
|
| 169 |
+
|
| 170 |
+
> **Note:** Update `BASE_DIR` in the script to match your local environment.
|
| 171 |
+
|
| 172 |
+
---
|
| 173 |
+
|
| 174 |
+
## Reliability & Validity
|
| 175 |
+
|
| 176 |
+
- **Inter-run consistency:** Each task evaluated 3 times; average used to reduce VLM variance
|
| 177 |
+
- **Prompt robustness:** JSON-only output enforced; fallback parsing with regex extraction
|
| 178 |
+
- **Rate limit handling:** 1.5s sleep between API calls
|
| 179 |
+
- **Token safety:** `max_tokens=1024` to prevent JSON truncation
|
tasks.json
ADDED
|
@@ -0,0 +1,247 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"name": "GIS-Bench",
|
| 3 |
+
"version": "1.0",
|
| 4 |
+
"description": "LLM 기반 GIS 자동화 에이전트 평가 벤치마크 — 45개 태스크",
|
| 5 |
+
"total_tasks": 45,
|
| 6 |
+
"levels": {
|
| 7 |
+
"Level1": {
|
| 8 |
+
"count": 16,
|
| 9 |
+
"description": "기본 시각화"
|
| 10 |
+
},
|
| 11 |
+
"Level2": {
|
| 12 |
+
"count": 17,
|
| 13 |
+
"description": "중급 공간 분석"
|
| 14 |
+
},
|
| 15 |
+
"Level3": {
|
| 16 |
+
"count": 12,
|
| 17 |
+
"description": "고급 / MCP 한계 탐색"
|
| 18 |
+
}
|
| 19 |
+
},
|
| 20 |
+
"tasks": [
|
| 21 |
+
{
|
| 22 |
+
"task_no": 1,
|
| 23 |
+
"level": 1,
|
| 24 |
+
"prompt": "서울시 대여소 파일을 불러와서 지도 위에 점으로 다 찍어줘."
|
| 25 |
+
},
|
| 26 |
+
{
|
| 27 |
+
"task_no": 2,
|
| 28 |
+
"level": 1,
|
| 29 |
+
"prompt": "배경에 실제 거리가 보이는 지도를 깔아줄 수 있어?"
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"task_no": 3,
|
| 33 |
+
"level": 1,
|
| 34 |
+
"prompt": "점의 색깔을 내가 보기 편하게 파란색으로 바꿔줘."
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"task_no": 4,
|
| 38 |
+
"level": 1,
|
| 39 |
+
"prompt": "도봉구에 있는 대여소만 지도에 남기고 나머지는 지워줘."
|
| 40 |
+
},
|
| 41 |
+
{
|
| 42 |
+
"task_no": 5,
|
| 43 |
+
"level": 1,
|
| 44 |
+
"prompt": "강북구에 위치한 CCTV만 점 위에 안심주소가 나오게 해줘."
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"task_no": 6,
|
| 48 |
+
"level": 1,
|
| 49 |
+
"prompt": "글자가 잘 안 보이는데, 글자 주변에 흰색 테두리를 넣어줘."
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"task_no": 7,
|
| 53 |
+
"level": 1,
|
| 54 |
+
"prompt": "대여소 점의 크기를 지금보다 2배로 키워줘."
|
| 55 |
+
},
|
| 56 |
+
{
|
| 57 |
+
"task_no": 8,
|
| 58 |
+
"level": 1,
|
| 59 |
+
"prompt": "특정 좌표(37.6398, 127.0255)를 중심으로 지도를 확대해줘."
|
| 60 |
+
},
|
| 61 |
+
{
|
| 62 |
+
"task_no": 9,
|
| 63 |
+
"level": 1,
|
| 64 |
+
"prompt": "현재 화면에 보이는 지도를 이미지 파일로 저장해서 보내줘."
|
| 65 |
+
},
|
| 66 |
+
{
|
| 67 |
+
"task_no": 10,
|
| 68 |
+
"level": 1,
|
| 69 |
+
"prompt": "데이터에 안심 주소가 창5S431인 곳만 노란색 별 모양으로 표시해줘"
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"task_no": 11,
|
| 73 |
+
"level": 1,
|
| 74 |
+
"prompt": "데이터에 주차 대수가 10대 넘는 곳만 노란색 별 모양으로 표시해줘"
|
| 75 |
+
},
|
| 76 |
+
{
|
| 77 |
+
"task_no": 12,
|
| 78 |
+
"level": 1,
|
| 79 |
+
"prompt": "지도 배경을 위성 모드 테마로 변경해줘"
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"task_no": 13,
|
| 83 |
+
"level": 1,
|
| 84 |
+
"prompt": "서울시 shp 파일이야. 그려서 보여줘."
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"task_no": 14,
|
| 88 |
+
"level": 1,
|
| 89 |
+
"prompt": "서울시 용도지구(고도지구) shp 파일이야. 그려서 보여줘."
|
| 90 |
+
},
|
| 91 |
+
{
|
| 92 |
+
"task_no": 15,
|
| 93 |
+
"level": 1,
|
| 94 |
+
"prompt": "DG,_AR에 따라 다르게 색상을 칠해서 보여줘"
|
| 95 |
+
},
|
| 96 |
+
{
|
| 97 |
+
"task_no": 16,
|
| 98 |
+
"level": 1,
|
| 99 |
+
"prompt": "DG,_AR가 북한산 주변인 것만 표시해줘"
|
| 100 |
+
},
|
| 101 |
+
{
|
| 102 |
+
"task_no": 17,
|
| 103 |
+
"level": 2,
|
| 104 |
+
"prompt": "서울시 CCTV가 뭉쳐있는 곳을 열지도(Heatmap)로 그려서 보내줘."
|
| 105 |
+
},
|
| 106 |
+
{
|
| 107 |
+
"task_no": 18,
|
| 108 |
+
"level": 2,
|
| 109 |
+
"prompt": "거치대수가 많은 곳은 큰 원으로, 적은 곳은 작은 원으로 그려줘."
|
| 110 |
+
},
|
| 111 |
+
{
|
| 112 |
+
"task_no": 19,
|
| 113 |
+
"level": 2,
|
| 114 |
+
"prompt": "열지도의 반경을 넓게 해서 더 부드럽게 퍼지게 그려줘."
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"task_no": 20,
|
| 118 |
+
"level": 2,
|
| 119 |
+
"prompt": "구 경계선을 지도 위에 굵은 검은색 선으로 표시해줘. (약 1분 정도 소요)"
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"task_no": 21,
|
| 123 |
+
"level": 2,
|
| 124 |
+
"prompt": "열지도 뒤에 배경 지도가 잘 보이게 히트맵의 불투명도를 조절해줘."
|
| 125 |
+
},
|
| 126 |
+
{
|
| 127 |
+
"task_no": 22,
|
| 128 |
+
"level": 2,
|
| 129 |
+
"prompt": "OpenStreetMap 배경 위에 고도지구 그라데이션을 겹쳐주고 그 위에 cctv 데이터를 겹쳐서 표현해줘"
|
| 130 |
+
},
|
| 131 |
+
{
|
| 132 |
+
"task_no": 23,
|
| 133 |
+
"level": 2,
|
| 134 |
+
"prompt": "제공된 UPIS_C_UQ123 SHP 파일 세트를 불러와서 지도에 표시해줘. 이때 누락된 파일이 있는지 확인하고 좌표계(CRS)가 무엇인지 알려줘."
|
| 135 |
+
},
|
| 136 |
+
{
|
| 137 |
+
"task_no": 24,
|
| 138 |
+
"level": 2,
|
| 139 |
+
"prompt": "구역별 CCTV 설치 밀도를 계산해서, 밀도가 낮은 구역은 빨간색, 높은 구역은 초록색으로 칠한 단계구분도(Choropleth Map)를 그려줘. 배경에는 실제 지도를 깔아줘."
|
| 140 |
+
},
|
| 141 |
+
{
|
| 142 |
+
"task_no": 25,
|
| 143 |
+
"level": 2,
|
| 144 |
+
"prompt": "CCTV 연계현황 CSV파일에서 안심주소 [3020][방범-번1-089](고정,북서)에서 걸어서 5분 거리(약 300m) 영역을 지도에 색칠해줘."
|
| 145 |
+
},
|
| 146 |
+
{
|
| 147 |
+
"task_no": 26,
|
| 148 |
+
"level": 2,
|
| 149 |
+
"prompt": "DRAWING_NO 숫자에 따라 그라데이션으로 색상을 표시해줘. 색상은 숫자가 적으면 흰색, 숫자가 높을수록 빨간색으로"
|
| 150 |
+
},
|
| 151 |
+
{
|
| 152 |
+
"task_no": 27,
|
| 153 |
+
"level": 2,
|
| 154 |
+
"prompt": "서울시 용도지구(고도지구) shp 파일이야. cctv csv 파일에 레이어드해서 같이 보여줄 수 있어?"
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"task_no": 28,
|
| 158 |
+
"level": 2,
|
| 159 |
+
"prompt": "CH_D010_00_20260221.shp 레이어에서 A9�� 따라 다른 색상으로 점을 표시해줘"
|
| 160 |
+
},
|
| 161 |
+
{
|
| 162 |
+
"task_no": 29,
|
| 163 |
+
"level": 2,
|
| 164 |
+
"prompt": "서울시에 위치한 의료시설만 남겨주세요."
|
| 165 |
+
},
|
| 166 |
+
{
|
| 167 |
+
"task_no": 30,
|
| 168 |
+
"level": 2,
|
| 169 |
+
"prompt": "cctv 위치가 500m 범위(버퍼)가 서로 겹치는 구간만 추출해서 보여줘"
|
| 170 |
+
},
|
| 171 |
+
{
|
| 172 |
+
"task_no": 31,
|
| 173 |
+
"level": 2,
|
| 174 |
+
"prompt": "DG,_AR가 북한산 주변인 서울시 용도지구(고도지구) 데이터를 가져와서 근처 100m 안에 있는 cctv를 찾아줘."
|
| 175 |
+
},
|
| 176 |
+
{
|
| 177 |
+
"task_no": 32,
|
| 178 |
+
"level": 2,
|
| 179 |
+
"prompt": "각 자치구 면적에 비해 CCTV가 너무 적은 서비스 취약구를 찾아줘."
|
| 180 |
+
},
|
| 181 |
+
{
|
| 182 |
+
"task_no": 33,
|
| 183 |
+
"level": 2,
|
| 184 |
+
"prompt": "cctv 데이터를 분석해서 서비스가 가장 필요한 지역을 선정하고, 그 이유를 설명하는 종합 지도로 보여줘"
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"task_no": 34,
|
| 188 |
+
"level": 3,
|
| 189 |
+
"prompt": "데이터를 분석 정확도를 높이기 위해 native:fillberry 또는 saga:fillsinks 알고리즘을 사용해 DEM의 결측치와 왜곡을 보정해주고 QGIS에서 보여주세요"
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"task_no": 35,
|
| 193 |
+
"level": 3,
|
| 194 |
+
"prompt": "DEM 파일을 QGIS 레이어로 추가해주고 native:slope 알고리즘을 사용하여 경사도(Slope) 레이어를 QGIS에서 생성해줘"
|
| 195 |
+
},
|
| 196 |
+
{
|
| 197 |
+
"task_no": 36,
|
| 198 |
+
"level": 3,
|
| 199 |
+
"prompt": "이 데이터는 전국 단위이므로, native:heatmapkerneldensity 알고리즘을 사용해서 산사태 발생 밀도를 나타내는 QGIS에 열지도(Heatmap)를 생성해줘요"
|
| 200 |
+
},
|
| 201 |
+
{
|
| 202 |
+
"task_no": 37,
|
| 203 |
+
"level": 3,
|
| 204 |
+
"prompt": "내 컴퓨터 대학원 졸업논문 폴더 안 sample data 폴더에서 강북구_안심이_CCTV 파일과 UPIS_C_UQ123 SHP 파일을 찾아서 레이어드해서 보여줘"
|
| 205 |
+
},
|
| 206 |
+
{
|
| 207 |
+
"task_no": 38,
|
| 208 |
+
"level": 3,
|
| 209 |
+
"prompt": "내 컴퓨터 대학원 졸업논문 폴더 안 sample data 폴더에서 한글 깨짐(utf-8/cp949)이 발생하는 파일을 찾아서 자동으로 인코딩을 수정한 뒤 지도를 그려줘"
|
| 210 |
+
},
|
| 211 |
+
{
|
| 212 |
+
"task_no": 39,
|
| 213 |
+
"level": 3,
|
| 214 |
+
"prompt": "서울 열린데이터광장에서 서울시 안심이 CCTV 연계 현황 검색해서 가장 최신 CSV 파일을 내려받아 지도에 나타내줘"
|
| 215 |
+
},
|
| 216 |
+
{
|
| 217 |
+
"task_no": 40,
|
| 218 |
+
"level": 3,
|
| 219 |
+
"prompt": "공공데이터포털에서 서울시 자치구별 인구 통계 데이터를 가져와서, 지금 로드된 CCTV 레이어와 결합해 인구 대비 CCTV 밀도 지도를 완성해줘"
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"task_no": 41,
|
| 223 |
+
"level": 3,
|
| 224 |
+
"prompt": "외국인 관광객을 위해 서울시 주요 관광지 SHP 데이터와 CCTV 데이터를 합친 뒤, 모든 라벨을 영어로 번역해서 지도를 만들어줘"
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"task_no": 42,
|
| 228 |
+
"level": 3,
|
| 229 |
+
"prompt": "인터넷에서 최근 1년간 서울시 범죄 발생 통계 기사를 찾아보고, UPIS_C_UQ123 구역 데이터와 비교해서 CCTV가 추가로 설치되어야 할 1순위 지점을 추천해줘"
|
| 230 |
+
},
|
| 231 |
+
{
|
| 232 |
+
"task_no": 43,
|
| 233 |
+
"level": 3,
|
| 234 |
+
"prompt": "C:/Users/A/Downloads/(B080)공개DEM_34602_img_2025/34602 데이터를 분석 정확도를 높이기 위해 native:fillberry 또는 saga:fillsinks 알고리즘을 사용해 DEM의 결측치와 왜곡을 보정해주고 QGIS에서 보여주세요"
|
| 235 |
+
},
|
| 236 |
+
{
|
| 237 |
+
"task_no": 44,
|
| 238 |
+
"level": 3,
|
| 239 |
+
"prompt": "C:/Users/A/Downloads/Landsat_위성영상의_정규물지수_2024 이 데이터는 전국 단위이므로, native:heatmapkerneldensity 알고리즘을 사용해서 산사태 발생 밀도를 나타내는 QGIS에 열지도(Heatmap)를 생성해줘요"
|
| 240 |
+
},
|
| 241 |
+
{
|
| 242 |
+
"task_no": 45,
|
| 243 |
+
"level": 3,
|
| 244 |
+
"prompt": "보정완료_DEM 레이어에 단일 밴드 가상 색채(Singleband Pseudocolor) 렌더러를 적용해줘. 색상 램프는 Magma 또는 Spectral을 사용하되, 고도 값을 5개 구간으로 나누어줘. 특히 해수면과 가까운 고도 0m~10m 사이는 짙은 파란색으로, 그 이상은 지형색으로 분류하여 저지대가 도드라져 보이게 스타일을 설정해줘"
|
| 245 |
+
}
|
| 246 |
+
]
|
| 247 |
+
}
|