| library_name: pytorch | |
| tags: | |
| - reinforcement-learning | |
| - board-game | |
| # GIPF Zero checkpoint | |
| Custom PyTorch `state_dict` checkpoint for the [GIPF Zero source](https://github.com/wusche1/gipf-zero). Load it with `training.model.load_model`. | |
| Published public opponent: **GIPF Zero — 43,869 games**, trained for **43869 self-play games**. Architecture: `{"blocks": 2, "head": "flat", "kind": "resnet", "width": 32}`. | |
| Source training checkpoint SHA-256: `aa533bb8922231a3e0dc53dd2ff947cab6899ef09a9245463c89ac99ea776ef0`. | |
| Served `champion.pt` artifact SHA-256: `d5a7e3bda2966a79d22cf34f131976907260e679683d0e2680d935312f41c2f6`. | |
| `RESULTS.md` documents the controlled overnight comparison of MLP, square-CNN, hex-CNN, and query-Transformer policies, plus its separately labelled continuation evaluations. These automated results are not a human or expert rating. | |
| [Full overnight comparison and results](RESULTS.md). | |