iviazovetskyi/enclosure-v2

Enclosure v2

Policy and value network for Enclosure, trained by self-play.

Code: https://github.com/iviazovetskyi/EnclosurePublic

Self-play games: https://huggingface.co/datasets/iviazovetskyi/enclosure-public

model.pt is a PyTorch checkpoint with keys cfg, model, and step. The optimizer state is not included.

• Architecture: 10 residual blocks, 128 channels, 4 attention blocks, about 4.2 million parameters (version 2).
• Training step: 500000.

v2-s24.wasm is the Constellation bot for this network: int8 weights, 24 playouts, built from this checkpoint.

Load the checkpoint with the bot server:

python rl/server.py

The server downloads this file on startup. A local .pt path also works.

License

MIT. See the GitHub repository.