SWE-Game Datasets

A collection of web-based mini-games used to build a software-engineering benchmark for coding agents. The core task is: given a small, self-contained game project, write Playwright end-to-end tests for it.

This repository ships three components of the data pipeline so that downstream users can reproduce filtering, re-label, or extend the benchmark.

Dataset Structure

swe-game/
├── README.md
├── quality_results.json          # 316 manual quality labels (high / low)
├── raw_games/                    # 1,069 raw mini-game projects (HTML / JS / CSS)
│   ├── [1000]吃豆豆/
│   ├── [1001]各种测试/
│   └── ...
└── benchmark_91/                 # 91 curated samples, each with a Playwright task
    ├── [1008]小鸟飞飞飞/
    │   ├── xiaoniaofeifei/       # game source code
    │   └── task.md               # task description (write 10 Playwright tests)
    └── ...

Components

1. raw_games/ — Raw Game Pool (1,069 games, ~6.7 GB)

The unfiltered pool. Each directory is a self-contained web project (usually an index.html plus static assets). Folder name pattern is [ID]<chinese_name>.

2. quality_results.json — Manual Quality Labels (316 entries)

Labels assigned by a human annotator who played each game in a browser:

Label	Count
high	91
low	225
Total labeled	316

Games not in this JSON are unlabeled (the annotation run stopped at 316 / ~836 candidates after a first auto prefilter).

{
  "[1000]吃豆豆": "low",
  "[1008]小鸟飞飞飞": "high",
  ...
}

3. benchmark_91/ — Curated Benchmark (91 samples)

All games labeled high in quality_results.json, each augmented with a task.md that defines a Playwright testing task (write exactly 10 test() cases covering the game's core loop).

This is the set used in the main experiments.

Data Pipeline

raw crawl                             quality_results.json
┌───────────────┐   auto prefilter   ┌───────────────┐   manual annotation   ┌───────────────┐
│ raw_games     │ ─────────────────▶ │ 836 candidates│ ───────────────────▶  │ 316 labels    │
│ 1,069 games   │                    │ (intermediate,│                       │ high=91       │
│               │                    │  not shipped) │                       │ low=225       │
└───────────────┘                    └───────────────┘                       └───────────────┘
                                                                                      │
                                                                                      │ keep `high`
                                                                                      ▼
                                                                              ┌───────────────┐
                                                                              │ benchmark_91  │
                                                                              │ + task.md     │
                                                                              └───────────────┘

Reproducing the Annotation

The interactive labeling server that produced quality_results.json is available in the companion GitHub repo:

https://github.com/YuyaoGe/swe-game-datasets → scripts/stage0/game_quality_filter.py

Usage:

# Label a folder of games (default args match the original run)
python scripts/stage0/game_quality_filter.py \
    --game-dir   ./raw_games \
    --result-file ./my_labels.json \
    --port       8765

# Then open http://localhost:8765 in a browser.
# Keys: 1/L = low, 2/H = high, S = skip, ←/→ = nav

Companion Code

Resource	Link
Experiment & benchmark code	https://github.com/YuyaoGe/swe-game-datasets

License

Apache 2.0.

Citation

If this dataset is useful to you, please cite the companion repository (see the GitHub README for the preferred BibTeX once the paper is released).