Datasets:

unhappyNN
/

agentic_forge

This repository releases the annotation tools and five-axis evaluation framework from our paper. The full benchmark comprises 29 apps, 309 screens, and 701 navigation test specifications (see paper).

Data release (preview). The dataset (input/mobileforge/) and experimental run artifacts (output/, experiments/) are provided as partial samples only, so you can inspect the format and try the evaluation pipeline. ~~The full release (all 29 apps, annotations, and 174 paper runs) will be published after the paper is accepted. Watch this repository for updates.~~(Datasets and experimental results: https://huggingface.co/datasets/unhappyNN/agentic_forge/)

Overview

Most design-to-code benchmarks focus on single pages and cannot assess multi-screen mobile products. MobileForge defines the task as:

Given multiple design screenshots of a real mobile app plus a page-relationship description, evaluate the generated frontend project on build success, navigation correctness, visual fidelity, code maintainability, and efficiency.

Gap	MobileForge response
Single-page granularity	Project-level: each instance is a folder of screens
No interactive navigation eval	701 navigation specs + state-isolated Playwright execution
No project-wide maintainability	LoC, component reuse, dead components, design-token consistency

Benchmark statistics

Metric	Value
Real consumer apps	29 (14 category buckets)
Human-reviewed screens	309 (mean 10.7 / app, range 5–21)
Navigation test specs	701 (mean 24.2 / app, range 11–41)
Page-relationship descriptions	One natural-language spec per app (tabs + parent–child)

Test types: 120 page-existence, 392 tab-navigation, 189 parent–child navigation.

The paper reports 174 end-to-end runs (6 models × 29 apps). Only a subset of example runs is included in this preview; the complete run bundle will be released upon acceptance.

Five-axis evaluation

Axis	Metrics	Module
Build	`tsc` + `vite build`; route count, page coverage	`evaluator` → `build`
Navigation	`NavPassRate` over 701 specs (state-isolated)	`testgen` → `navigation_test`
Visual fidelity	Anchor-reference list-wise Borda (primary)	`testgen` → `navigation_visual`
Code maintainability	LoC, ReuseRate, DeadCompRate, ColorConsistency, …	`evaluator` → `project_quality`
Efficiency	Tokens, call count, wall time, cost	`evaluator` → `efficiency`

State-isolated navigation testing. Each case starts from a fixed source route (e.g. /search), not a chained walk from home, so early routing bugs do not cascade into later failures.

Anchor-reference list-wise visual judging. The reference screenshot is inserted as a hidden anchor; rankings that do not place the anchor first are discarded. The paper uses Gemini 2.5 Pro as the fixed judge.

Repository layout

MobileForge/
├── input/mobileforge/          # Partial sample apps (full set after acceptance)
├── output/                     # Partial sample runs (reference only)
├── experiments/                # Partial sample experiment dirs (reference only)
├── annotator/                  # Annotation pipeline + check-tool/ review UI
├── testgen/                    # Navigation tests + visual judging
├── evaluator/                  # Five-axis evaluation + reports
├── experiment.yaml
├── models.yaml
├── env.sh.example
└── requirements.txt

Module docs:

[annotator/README.md](annotator/README.md)
[annotator/check-tool/README.md](annotator/check-tool/README.md)
[testgen/README.md](testgen/README.md)
[evaluator/README.md](evaluator/README.md)

Quick start

pip install -r requirements.txt
python -m playwright install chromium
nvm use 20

cp env.sh.example env.sh && source env.sh
# Edit experiment.yaml (see experiment.yaml.copy)

Use the sample apps under input/mobileforge/ (partial preview). Each sample app needs metadata.json and test_cases.json. You may also point the evaluator at your own generated projects.

Evaluate one generated project (must be a buildable React + TypeScript + Tailwind tree):

python -m evaluator eval \
  --dir <path-to-generated-project> \
  --input ./input/mobileforge/airbnb

Batch evaluation:

python -m evaluator eval --dir ./output/mobileforge \
  --input ./input/mobileforge --batch

python -m evaluator report --experiment-dir experiments/<timestamp>_<name>

Pipeline

flowchart LR
  A[Screenshots] --> B[Annotator]
  B --> C[metadata + test_cases]
  C --> D[Generated project]
  D --> E[Evaluator / TestGen]
  C --> E

Annotation

python -m annotator --input ./input/mobileforge/airbnb
cd annotator/check-tool && python server.py --input-dir ../../input/mobileforge

metadata.json — page relationships and per-screen annotations (ground truth)
test_cases.json — navigation test specifications (ground truth)

Evaluation

python -m evaluator eval --dir <run_dir> --input ./input/mobileforge/<app>
python -m evaluator eval --dir <run_dir> --input ./input/mobileforge/<app> \
  --only navigation_test,navigation_visual
python -m evaluator eval --dir <run_dir> --input ./input/mobileforge/<app> \
  --skip navigation_visual

Standalone testgen:

python -m testgen run --cases input/mobileforge/<app>/test_cases.json \
  --run-dir <run_dir> --input-dir input/mobileforge/<app>
python -m testgen eval --report <run_dir>/eval/test_report.json \
  --cases input/mobileforge/<app>/test_cases.json \
  --input-dir input/mobileforge/<app>

Evaluation order: build → navigation_test → navigation_visual; other modules are independent.

Data format

input/mobileforge/<app>/
├── *.png / *.jpg
├── desc.txt              # optional
├── metadata.json
└── test_cases.json

Results are written under <run_dir>/eval/ (build_result.json, test_report.json, summary.json, …).

Comparison to prior benchmarks

Benchmark	Unit	Build	Nav.	Visual	Maint.	Eff.
Design2Code, WebSight, …	Single page	—	—	✓	—	—
Figma2Code	Figma files	—	—	✓	✓	—
MobileForge	29 apps / 309 screens	✓	✓	✓	✓	✓

Inputs are screenshots only (no Figma structural prior). Evaluation targets runnable multi-page web projects (React + TypeScript + Tailwind in the paper).

Ethics

Screenshots are captured from publicly distributed apps with no real-user PII. The benchmark is for research evaluation only; the release license prohibits commercial reuse or re-skinning of screenshot assets.

Citation

@inproceedings{mobileforge2026,
  title     = {Looks Right, Works Right: A Project-Level Benchmark for Multi-Screen Mobile App Generation},
  author    = {Anonymous},
  booktitle = {Proceedings of EMNLP},
  year      = {2026}
}

FAQ

Where is the full benchmark?
Not in this preview. Only partial samples are shipped for format reference and pipeline testing. All 29 apps, ground-truth annotations, and 174 paper runs will be released after acceptance.

Missing input/mobileforge/ or sample apps?
The preview may include only a few apps; large assets may be gitignored. Check the repository release notes. Full data will follow acceptance.

Navigation tests skipped?
Ensure the project builds, pass --input, confirm test_cases.json exists, and install Playwright (playwright install chromium).

Reproduce the paper’s 174 runs?
Not possible with the preview bundle alone. After the full release, run evaluator eval --batch on the published runs/ directory, then evaluator report.

High visual-judge API cost?
Use --skip navigation_visual; see testgen/visual_judge_listwise.py for list-wise details.

Downloads last month: 42