timmyburn/hexo-bootstrap-corpus

HeXO Bootstrap Corpus

Supervised pretraining corpus for seeligto/hexo_rl, an AlphaZero-style learner for Hex Tic-Tac-Toe (infinite hexagonal grid, 6-in-a-row).

Access: this dataset is currently private. Contact the repo owner for access. A sanitized, public-licensed release may follow later.

Contents

Single NPZ file with four aligned arrays:

Key	Shape	Dtype	Meaning
states	(N, 18, 19, 19)	float16	Board tensors (AlphaZero-style 18-plane encoding)
policies	(N, 362)	float32	Target move distributions over the 19×19 action grid + pass
outcomes	(N,)	float32	Game outcomes from the side-to-move: +1 win, -1 loss, 0 draw
weights	(N,)	float32	Per-sample training weights (Elo-band biased for human games)

N ≈ 320k positions across ~19k games.

Sources

• Anonymized public human games — visibility=public only; all PII (usernames, profile IDs, session IDs, exact timestamps) stripped at ingestion.
• SealBot self-play games — community minimax engine at mixed time limits (fast ≈ 0.05 s, strong ≈ 0.5 s per move).
• Injected hybrid games — human-seed openings continued by SealBot, used to broaden the opening distribution.

PII policy

The NPZ payload contains only board tensors and move labels — no textual or identifying metadata. Human games were processed through a one-way PII-scrubbing pipeline before corpus export: displayName, playerId, sessionId, and exact millisecond timestamps are removed. profileId values are replaced with deterministic 16-character SHA-256 hashes (enabling cross-game aggregation without exposing identity). Games flagged visibility: private are not included.

Loading

Memory-map to avoid loading the full 4.6 GB into RAM:

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="timmyburn/hexo-bootstrap-corpus",
    filename="bootstrap_corpus.npz",
    repo_type="dataset",
)
data = np.load(path, mmap_mode="r")
states, policies, outcomes, weights = (
    data["states"], data["policies"], data["outcomes"], data["weights"]
)
print(states.shape, states.dtype)

Authentication: run hf auth login once, or set HF_TOKEN=hf_xxx.

Companion model

timmyburn/hexo-bootstrap-models — policy/value net pretrained on this corpus (public).

Files

File	Size	Description
bootstrap_corpus.npz	~4.6 GB	Mixed-source supervised corpus (uncompressed NPZ, float16 states)

License

MIT — see the repository LICENSE.