evaluation/eval_criteria.md

GIS-Bench Evaluation Criteria

Overview

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).

• Judge Model: claude-sonnet-4-6
• Evaluation Repetitions: N = 3 per task (averaged for robustness)
• Raw Score Scale: 130 points
• Normalized Score Scale: 100 points
• Pass Threshold: 60 points (normalized)

---

Scoring Criteria

Category 1: Basic Quality (100 pts)

Criterion	Max Score	Scoring Rule
exists	20	0 or 20 — Meaningful visualization exists (not blank/error)
accuracy	30	0–30 — Requested region and dataset are correctly used
requirement	30	0–30 — All conditions met (color, filter, buffer, etc.)
completeness	20	0–20 — Visual completeness: legend, title, labels present

Category 2: Spatial Accuracy (30 pts)

Visual judgment from screenshot.

Criterion	Max Score	Scoring Rule
spatial_location	15	15 = correct location / 8 = minor offset / 0 = coordinate error or wrong area
geometry_validity	5	5 = normal / 2 = minor distortion / 0 = obvious distortion
numeric_match	10	10 = fully reflected / 5 = partially reflected / 0 = not reflected

---

Score Normalization

total_raw        = exists + accuracy + requirement + completeness
                   + spatial_location + geometry_validity + numeric_match

total_normalized = round(total_raw / 130 * 100)

pass             = (total_normalized >= 60)

---

VLM Judge Prompt

The following prompt is passed to the judge model along with the task screenshot (base64 image):

GIS 결과물 평가 전문가로서 아래 태스크의 출력 이미지를 채점하세요.

[태스크 Level {level}]
{prompt}

[채점 기준 — 130점 만점]
기본(100점):
  exists(0/20): 의미있는 시각화 존재 여부
  accuracy(0-30): 요청 지역·데이터셋 정확성
  requirement(0-30): 색상·필터·버퍼 등 조건 충족
  completeness(0-20): 범례·제목·라벨 완성도

공간정확성(30점, 스크린샷 시각 판단):
  spatial_location(0/8/15): 데이터가 올바른 지역에 표시되는지
    15=정상, 8=경미한 이상, 0=좌표오류·엉뚱한위치
  geometry_validity(0/2/5): 폴리곤·버퍼 형태 이상 여부
    5=정상, 2=경미한이상, 0=명백한왜곡
  numeric_match(0/5/10): 수치·조건이 결과에 반영되었는지
    10=일치, 5=부분반영, 0=미반영

total_raw = 7개 항목 합계
total_normalized = round(total_raw / 130 * 100)

[응답] JSON만 출력, 다른 텍스트 금지:
{"exists":정수,"accuracy":정수,"requirement":정수,"completeness":정수,
 "spatial_location":정수,"geometry_validity":정수,"numeric_match":정수,
 "total_raw":정수,"total_normalized":정수,"pass":불리언,"reason":"한줄이유"}

---

Evaluation Flow

Output File
    │
    ├── .html / .htm  →  Playwright rendering (3s wait)  →  Screenshot (PNG)
    │                                                              │
    └── .png / .jpg   ────────────────────────────────────────────┘
                                                                   │
                                                          Base64 encode
                                                                   │
                                                     Claude VLM Judge × 3
                                                                   │
                                                        Average scores
                                                                   │
                                                     Pass / Fail determination

---

Output Fields

Each evaluated task produces the following fields:

Field	Type	Description
task_no	int	Task ID
level	str	Level1 / Level2 / Level3
model	str	Model identifier
file_type	str	html_screenshot / png / jpg
score	float	Average normalized score (0–100)
score_raw	float	Average raw score (0–130)
spatial_location	float	Average spatial location score
geometry_validity	float	Average geometry validity score
numeric_match	float	Average numeric match score
pass	bool	True if score ≥ 60
reason	str	VLM reason summary
scores_3x	str	Individual normalized scores from 3 runs
scores_3x_raw	str	Individual raw scores from 3 runs

---

Evaluation Script

The full evaluation pipeline is available at: evaluation/run_evaluation_vlm_v3.py

Requirements

anthropic
playwright
openpyxl
pandas

Usage

# Set API key
set ANTHROPIC_API_KEY=your_key_here   # Windows
export ANTHROPIC_API_KEY=your_key_here  # Linux/Mac

# Install Playwright browser
playwright install chromium

# Run evaluation
python run_evaluation_vlm_v3.py

Directory Structure Expected

C:\Users\A\Desktop\대학원 졸업논문\
├── GIS_Bench_문항.xlsx          # Task prompts (Excel)
├── claude_mcp\                  # Model output files
│   ├── 1.html
│   ├── 2.png
│   └── ...
├── claude_baseline\
└── gpt_5.2\

Note: Update BASE_DIR in the script to match your local environment.

---

Reliability & Validity

• Inter-run consistency: Each task evaluated 3 times; average used to reduce VLM variance
• Prompt robustness: JSON-only output enforced; fallback parsing with regex extraction
• Rate limit handling: 1.5s sleep between API calls
• Token safety: max_tokens=1024 to prevent JSON truncation