The dataset viewer is not available for this subset.
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Smash or Transformer -- mixed_v1 dataset
The training dataset for the recommended Smash or Transformer model: 21,076 images across all 1,025 Pokemon (official artwork + in-game sprites + Safebooru fan-art), with crowd "smash" labels from pokesmash.xyz.
Code, docs, and reproduction guide: https://github.com/byrte1024/SmashOrTransformer/blob/master/docs/vit_small_mixed_v1.md
Contents
mixed_v1_dataset.tar (~11 GB) extracts at the repo root to:
datasets/mixed_v1/-- packed image blob (images.bin) + metadata (data.npz,split.json,manifest.csv,stats.json,config.json)backgrounds/-- real + battle backdrops the sprite augmentation composites onto
Usage
git clone https://github.com/byrte1024/SmashOrTransformer && cd SmashOrTransformer
uv sync
uv run python download_models.py --dataset # downloads + extracts this tar
uv run python -m model.train configs/train_mixed.json
License
other -- this bundles scraped Safebooru fan-art and PokeAPI / official sprite
assets. These are third-party works redistributed for research and
reproducibility, with attribution to the original artists, PokeAPI, and
Nintendo. Not for commercial use; not ours to relicense.
- Downloads last month
- 29