Atomic-Germ/Pygame-data

---
license: apache-2.0
language:
- en
tags:
- code-generation
- pygame
- game-development
- hyper-casual
- python
- text-to-code
size_categories:
- 1K<n<10K
---

# Pygame Hyper-Casual Game Generation Dataset

A supervised fine-tuning dataset for training language models to generate complete, runnable [Pygame](https://www.pygame.org/) games from natural language prompts. Each entry pairs a plain-English game description with a self-contained Python script that implements the described game.

Originally created for training **89-Times**, an AI game designer specialized in hyper-casual mobile games.

## Dataset Summary

| | |
|---|---|
| **Total examples** | 2,206 |
| **Unique game prompts** | 200 |
| **Unique game concepts** | 182 |
| **Language** | English |
| **Code framework** | Pygame (Python) |
| **Target platform** | Mobile portrait (touch/click input) |
| **Avg. code length** | ~13 lines per response |
| **License** | Apache 2.0 |

## Format

Each entry is a JSON object with two fields:

```json
{
"question": "Create a game where I have to click on a moving red balloon to score points.",
"response": "import pygame\nimport random\npygame.init()\n..."
}
```

- **`question`** — A natural language prompt describing a hyper-casual game mechanic.
- **`response`** — A complete, runnable Python script using Pygame that implements the game.

> **Note:** The response field was originally named `reponse` but was fixed to `response`.

## Files

| File | Entries | Description |
|------|---------|-------------|
| `train.json` | 2,000 | Main training set — prompt/code pairs across all 200 game types, with 10 code variations per prompt |
| `gamiasdf.json` | 6 | System identity and boundary examples (refusal responses for out-of-scope requests like 3D or multiplayer games) |
| `hsj.json` | 20 | Game prompt/code pairs |
| `lila.json` | 30 | Game prompt/code pairs |
| `loapz.json` | 30 | Game prompt/code pairs |
| `maya.json` | 30 | Game prompt/code pairs |
| `nia.json` | 40 | Game prompt/code pairs |
| `sara.json` | 30 | Game prompt/code pairs |
| `valentin.json` | 30 | Game prompt/code pairs |

The smaller files (`hsj` through `valentin`) each contain a unique code implementation for each game prompt. `train.json` contains 10 distinct code variations per prompt, providing diversity for training.

## Game Types

The dataset covers **182 unique hyper-casual game concepts** spanning a variety of mechanics:

- **Tap/Click** — Whack-a-Mole, Tap Speed, Quick Tap, Balloon Pop, Pop It, Piano Tiles
- **Timing** — Door Slam, Hammer Strength, Perfect Slice, Coffee Pour, Trampoline Jump
- **Drag/Slide** — Parking, Car Wash, Ice Skating, Soap Slide, Lawn Mower, Fruit Peel
- **Aim & Launch** — Slingshot, Basketball, Golf, Bowling, Sling Shot, Knife Throw
- **Dodge & Avoid** — Dodge, Space Dodge, Laser Avoid, Neon Dash, Traffic Lane, Hurdles
- **Balance & Physics** — Balancing, Gravity Flip, Balance Bar, Balloon Balance, Bungee Jump
- **Hold & Release** — Fishing, Oxygen Dive, Drilling, Rocket Landing, Soap Bubble Blow
- **Puzzle/Match** — Color Switch, Shape Match, Safe Cracker, Plumber, Match Pair, Lock Pick

## Code Characteristics

- **Self-contained** — Every response is a complete script that runs with just `import pygame`.
- **Screen resolutions** — Mobile-oriented portrait layouts: `360×640`, `400×800`, `300×600`, `720×1280`.
- **Input model** — Mouse/touch events (`MOUSEBUTTONDOWN`, `MOUSEMOTION`, `MOUSEBUTTONUP`), designed for single-touch hyper-casual gameplay.
- **Compact** — Games average ~13 lines, emphasizing single-mechanic simplicity.
- **No external assets** — All visuals use Pygame primitives (circles, rectangles, lines, polygons).

## Intended Use

- **Fine-tuning** code generation models to produce Pygame games from natural language descriptions.
- **Few-shot prompting** for game generation tasks.
- **Studying** patterns in minimal hyper-casual game implementations.

## Limitations

- Games are intentionally minimal — single-mechanic prototypes, not polished products.
- No sound, no sprite assets, no multi-level progression.
- Input is mouse-based (simulating mobile touch); no keyboard controls.
- Some game loops lack proper exit handling or scoring beyond `print()`.
- The `response` field name is a legacy artifact and is not corrected to preserve compatibility.

## Origin

This dataset was originally authored in French and machine-translated to English. The Python code was preserved as-is, with only embedded French strings (print statements, comments) translated.