timmyburn/hexo-bootstrap-models

HeXO Bootstrap Model

Pretrained policy/value network for Hex Tic-Tac-Toe — 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.

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.

Usage

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 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.