Datasets:

DylanRiden
/

smb-worldmodel-data

license: mit
task_categories:
  - video-classification
  - reinforcement-learning
tags:
  - world-model
  - nes
  - super-mario-bros
  - game-ai
size_categories:
  - 100K<n<1M

SMB World Model Training Data

Training data for a Super Mario Bros world model using the Titan memory architecture.

Dataset Description

118,166 frames from 8 TAS (Tool-Assisted Speedrun) playthroughs
Frame size: 224x256x3 (RGB)
Action space: 8 buttons [Up, Down, Left, Right, A, B, Start, Select]
Format: Compressed .npz files (each contains frame + action bundled together)
Total size: ~293MB (compressed)

TAS Files Used

TAS File	Description
smb_all_items	Collects all items
smb_low_percent	Minimal item collection
smb_max_coins	Maximum coins
smb_max_score	Maximum score
smb_min_a_presses	Minimal A button presses
smb_scoreless	Zero score run
smb_warpless	No warp zones
smb_warps	Using warp zones

Data Format

Each .npz file contains:

data = np.load("frame_000000.npz")
frame = data['frame']   # shape: (224, 256, 3), dtype: uint8
action = data['action'] # shape: (8,), dtype: float32

Action order: [Up, Down, Left, Right, A, B, Start, Select]

Usage

from huggingface_hub import hf_hub_download
import zipfile

# Download
path = hf_hub_download(
    repo_id="DylanRiden/smb-worldmodel-data",
    filename="smb_frames.zip",
    repo_type="dataset"
)

# Extract
with zipfile.ZipFile(path, 'r') as z:
    z.extractall("./nes_data")

Collection Method

Emulator: FCEUX with Lua scripting
Frame skip: Every 4th frame (15fps from 60fps)
Menu skip: First 250 frames skipped
Real-time conversion to bundled .npz format

License

MIT