---
license: mit
pretty_name: Hexo Human Corpus (encoding-free)
task_categories:
  - other
tags:
  - hex
  - hex-tac-toe
  - board-games
  - game-records
  - reinforcement-learning
  - alphazero
size_categories:
  - 1K<n<10K
---

# Hexo Human Corpus

Encoding-free corpus of **6,902** decisive human *Hex Tac Toe* games — hexagonal
grid, six-in-a-row to win (player 1 opens with 1 move, then both players play 2
moves per turn; the board is theoretically infinite).

Each line is one game as a **raw axial move list + outcome**. Nothing about any
neural-network encoding is baked in — no planes, no fixed board size, no action
space. Read it with the stdlib `json` module and build whatever representation
you want.

## Files

| file | description |
|------|-------------|
| `hexo_human_corpus.jsonl` | the corpus — one game per line |
| `SCHEMA.md` | full per-line schema + conventions |
| `dataset_metadata.json` | provenance: counts, sha256, source filter |

## Schema

One JSON object per line:

```json
{"game_hash":"0f8c6bdfc55e7f6f","moves":[[0,0],[2,-2],[-3,3]],"winner":1,"source":"human","elo":[898,955]}
```

| field | type | meaning |
|-------|------|---------|
| `game_hash` | string (16 hex) | SHA-256 of the move sequence — stable content/dedup key |
| `moves` | array of `[x, y]` | axial hex coords `(x,y)=(q,r)`, in play order |
| `winner` | `1` or `-1` | `1` = first player (X) wins, `-1` = second player (O) |
| `source` | string | `"human"` |
| `elo` | `[int\|null, int\|null]` | `[elo_p1, elo_p2]` |

**Conventions**

- Axial hex coordinates `(x, y)`; the board is infinite so values can be
  negative. The first player's forced opener is always `(0, 0)`.
- Replay `moves` in order to reconstruct any board state.
- Only decisive (six-in-a-row) games are included — there are **no draws**.

## Usage

```python
import json

games = [json.loads(line) for line in open("hexo_human_corpus.jsonl")]
print(len(games), "games")          # 6902
g = games[0]
print(g["moves"], g["winner"])      # [[0,0], [2,-2], ...]  1
```

## Provenance

Rated human games filtered to: rated, ≥20 moves, decisive by six-in-a-row.
Per-game `elo` is each player's rating at game time. Games are anonymised
(player ids dropped; only relative Elo retained). See `dataset_metadata.json`
for the exact sha256 and counts.

License: MIT.