blanchon/tinyworlds_parquet

TinyWorlds (parquet)

Retro game frames from AlmondGod/tinyworlds, repackaged into a uniform per-frame parquet schema with one split per game for fast, random-access frame loading.

TinyWorlds is a Genie reimplementation: it has no action labels and learns latent actions from video alone. These splits are therefore action-less -- ideal for training an image tokenizer (RAE/VAE) or an unconditional video world model.

Splits

split	resolution	~fps	actions
zelda	128x128	~15	none (latent)
pong	64x64	~30	none (latent)
sonic	64x64	~15	none (latent)
picodoom	64x64	~30	none (latent)
pole_position	64x64	~15	none (latent)

Schema

Each row is one frame, in playback order (a single contiguous episode per game):

• image — the RGB game frame (Image, PNG, lossless)
• episode — contiguous-play segment id (always 0 here; one continuous sequence per game)
• frame_index — frame position within the episode

Usage

from datasets import load_dataset
zelda = load_dataset("blanchon/tinyworlds_parquet", split="zelda")
zelda[0]["image"]  # PIL.Image

Derived from AlmondGod/tinyworlds (game footage downsampled from gameplay videos). Provenance and upstream license follow the original; redistributed here for research.