You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Arkadium

Arkadium Gin Rummy Dataset

High-Quality Human Gin Rummy Gameplay (Logs + Screenshots)
Matches   Screenshots   Size
Opening Hand
Opening Hand
Mid-Game
Mid-Game Decisions
Knock
Knock / Endgame

Dataset Overview

This dataset contains real human gameplay from Arkadium's gin rummy platform, providing researchers and developers with high-quality training data for game AI, reinforcement learning, and computer vision applications. All hands include screenshots.

RAW data only. The primary dataset is a single raw ZIP archive (.pbn/.webp). Parquet files in data/ contain only 10 example records for the HuggingFace Dataset Viewer and are not a full export. We do not provide a full Parquet version.

Totals (selected subset): 10,000 matches, 686,403 screenshots, ~23.62 GB compressed ZIP size. The dataset is capped at 10,000 unique hands with screenshots and packaged into a single archive.

Metric gamelogs_with_images
Raw match logs 10,000
Raw screenshots 686,403
Raw ZIP size ~23.62 GB
Raw format PBN + WebP (ZIP)

Use Cases

  • Game AI Training: Train models to predict optimal draws/discards using supervised learning on human gameplay.
  • Reinforcement Learning: Use as demonstration data for imitation learning or to initialize RL agents.
  • Computer Vision: Card/board state recognition, meld detection, and UI state extraction from images.
  • Strategy Analysis: Study human decision-making patterns, draw/knock timing, and deadwood management.

Data Format

File Structure

Arkadium Gin Rummy Dataset
│
└── Arkadium_Ginrummy_full_data_gamelogs_and_images_002.zip
    ├── 0130003.pbn
    ├── attachments_0130003/
    │   ├── 0130003_0001.webp
    │   ├── 0130003_0002.webp
    │   └── ...
    └── ...

Game Log Format (.pbn)

Each .pbn file contains a complete gin rummy hand record in a human-readable format:

[FF "4"]
[DT "2025.08.15"]
[GN "ginrummy"]
[PC "arkadium.com"]
[GameplayId "132069239"]

[Player1 "p1"]
[Player2 "p2"]

[Hand "3:6"]
[GameScore "63:0"]
[Dealer "2"]
[Deal "C2 S3 C5 D6 S6 H7 C8 C9 DQ SQ, HA DA C3 D4 H6 C6 S8 CJ SJ DK"]
[Upcard "HQ"]
[Score "0:25"]

[Play]
1. 1 picks HQx
2. 1 discards H7
3. 2 picks D2
4. 2 discards DK
...

Header Field Reference
Field Description Example
[DT] Hand date (YYYY.MM.DD) "2025.08.15"
[GN] Game name "ginrummy"
[PC] Platform "arkadium.com"
[GameplayId] Unique gameplay identifier "132069239"
[Player1] First player ID "p1"
[Player2] Second player ID "p2"
[Hand] Hand index as recorded (format X:Y) "3:6"
[GameScore] Cumulative score "63:0"
[Dealer] Dealer (1 or 2) "2"
[Deal] Initial hands (Player1, Player2) "C2 S3 ... , HA DA ..."
[Upcard] Initial upcard "HQ"
[Score] Hand score result "0:25"
[Deadwood] Deadwood totals "0:0"

Additional metadata tags may appear (e.g., [FF], [GM], [SZ], [CA], [RU]) and are preserved in the raw logs.

Play Notation Guide
Notation Meaning Example
picks XX Draws from the deck picks D2
picks XXx Draws from the discard pile picks S8x
discards XX Discards a card discards H7
passes Declines the upcard passes
knocks Knocks to end the hand knocks

Parquet Preview (10 records)
Parquet files in `data/` are **only a 10-record preview** for the HuggingFace Dataset Viewer. They exist **exclusively for demo/browsing** and **do not represent the full dataset**. All primary data is in the raw ZIP archives above, and no full Parquet release is provided.

Quick Start (Preview)

Installation

pip install datasets pillow

Load the Preview

from datasets import load_dataset

# Load preview with screenshots (10 records)
ds = load_dataset("ArkadiumInc/ArkadiumGinrummy", "gamelogs_with_images")

Basic Usage

# Access a single hand
hand = ds["train"][0]

print(f"Players: {hand['player1']} vs {hand['player2']}")
print(f"Date: {hand['event_date']}")
print(f"Score: {hand['score']}")
print(f"\nMoves:\n{hand['moves']}")

Working with Screenshots

from PIL import Image
import io

# Get a hand with images
hand = ds["train"][0]

print(f"This hand has {hand['num_images']} screenshots")

# Display the first screenshot
if hand['images']:
    img = Image.open(io.BytesIO(hand['images'][0]))
    img.show()

Parquet Schema (Preview)

These columns describe the Parquet preview (10 records).

Column Type Description Config
data_id string Unique record identifier gamelogs_with_images
file_id string Source file reference gamelogs_with_images
place string Platform gamelogs_with_images
num_images int32 Screenshot count gamelogs_with_images
images list[Image] WebP screenshots (stored as bytes/path in Parquet preview) gamelogs_with_images
player1 string First player ID gamelogs_with_images
player2 string Second player ID gamelogs_with_images
event_date string Hand date (YYYY.MM.DD) gamelogs_with_images
gameplay_id string Gameplay identifier gamelogs_with_images
hand string Hand score gamelogs_with_images
game_score string Cumulative score gamelogs_with_images
dealer string Dealer gamelogs_with_images
deal string Initial hands gamelogs_with_images
upcard string Upcard gamelogs_with_images
score string Hand score result gamelogs_with_images
moves string Parsed play sequence gamelogs_with_images
groups string Meld groupings gamelogs_with_images
layoffs string Layoffs by player gamelogs_with_images
deadwood_cards string Deadwood cards by player gamelogs_with_images
deadwood string Deadwood totals gamelogs_with_images
raw_log string Original .pbn content gamelogs_with_images

Citation

If you use this dataset in your research, please cite:

@dataset{arkadium_ginrummy_2026,
  title     = {Arkadium Gin Rummy Dataset: Human Gameplay with Screenshots},
  author    = {Arkadium Inc.},
  email     = {ai@arkadium.com},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/ArkadiumInc/ArkadiumGinrummy}
}
Downloads last month
68