File size: 2,884 Bytes
22de51c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 | ---
license: mit
library_name: pytorch
tags:
- reinforcement-learning
- alphazero
- board-games
- hex-tic-tac-toe
- mcts
pipeline_tag: other
---
# HeXO Bootstrap Model
Pretrained policy/value network for [Hex Tic-Tac-Toe](https://hex-tic-tac-toe.github.io/) —
a two-player game on an infinite hexagonal grid, 6-in-a-row to win. Used as
the starting point for AlphaZero-style self-play training in
[`seeligto/hexo_rl`](https://github.com/seeligto/hexo_rl).
## Architecture
- Input: **18 × 19 × 19** float tensor (AlphaZero-style history + scalar planes)
- ResNet-12 trunk with squeeze-and-excitation blocks
- GroupNorm(8) throughout (BN-free, stable under small batch sizes)
- Dual-pool value head with BCE loss
- Auxiliary heads: ownership prediction + winning-line prediction
- Saved as a `state_dict` inside a standard `torch.save` checkpoint
The board is genuinely infinite: the accompanying Rust engine uses a sparse
coordinate hashmap. The network receives a 19×19 window assembled around
the active stone cluster, so the model itself has no board-size prior.
## Training
Supervised bootstrap only — no self-play was used to produce this artifact.
Trained on a mixed corpus of:
- **SealBot self-play games** (community minimax engine, mixed time limits)
- **Anonymized public human games** (visibility=public, PII-stripped at ingestion)
- **Hybrid human-seed + bot-continuation games**
See the companion dataset (access-restricted):
[`timmyburn/hexo-bootstrap-corpus`](https://huggingface.co/datasets/timmyburn/hexo-bootstrap-corpus).
## Usage
```python
import torch
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="timmyburn/hexo-bootstrap-models",
filename="bootstrap_model.pt",
)
ckpt = torch.load(path, map_location="cpu", weights_only=False)
# Load into the network defined in seeligto/hexo_rl:
# from hexo_rl.model.network import HexTacToeNet
# model = HexTacToeNet(in_channels=18)
# model.load_state_dict(ckpt["model"])
# model.eval()
```
The full inference path (windowing, legal-move masking, policy projection
over the 362-dim action space) lives in the
[`hexo_rl`](https://github.com/seeligto/hexo_rl) repo.
## Evaluation
Calibrated against a threat-detection probe on 18-plane fixtures:
| Metric | Pass threshold | Notes |
|---|---|---|
| C2: extension cell in policy top-5 | ≥ 25% | baseline for bootstrap-v4 |
| C3: extension cell in policy top-10 | ≥ 40% | baseline for bootstrap-v4 |
Thresholds are minimum-viable — later self-play checkpoints should clear
these comfortably and will be released as a separate model variant.
## Files
| File | Size | Description |
|---|---|---|
| `bootstrap_model.pt` | ~17 MB | PyTorch checkpoint (state dict + optimizer + metadata) |
## License
MIT — see the [repository LICENSE](https://github.com/seeligto/hexo_rl/blob/master/LICENSE).
|