HexN Duel Corpus
105,000 two-player Catan games between six neural-network agents, P1 to P6. Each pair of agents played 5,000 games, and each agent played 5,000 games against itself, for 21 files in all.
We estimate each agent at full strength in the top 5% of human players; moves here are sampled at temperature 0.5 for variety.
Each game is a full record: the board, every action, and every dice roll and card draw. Games replay in HexSet 1.26.3.
P3 is catanrl-v2 by mcrco, used with permission.
CATAN is a trademark of Catan GmbH. This dataset is not affiliated with, endorsed by or sponsored by Catan GmbH or Catan Studio.
Rules
HexSet's duel-variant rules:
- two players, 15 victory points to win;
- a player with more than 9 cards discards half on a seven;
- the robber cannot be placed on a hex touching a building of a player with 2 or fewer public points;
- balanced dice (
hexset.chance.Balanced); - no trading between players (bank and harbor trades are allowed).
Results
Win rate of the row agent against the column agent, in percent, over 5,000 games (about ±1.4 points at 95%).
| P1 | P2 | P3 | P4 | P5 | P6 | |
|---|---|---|---|---|---|---|
| P1 | 63.4 | 63.4 | 62.4 | 68.4 | 69.6 | |
| P2 | 36.6 | 49.8 | 47.9 | 55.2 | 55.8 | |
| P3 | 36.6 | 50.2 | 48.1 | 55.3 | 54.7 | |
| P4 | 37.6 | 52.1 | 51.9 | 56.3 | 57.2 | |
| P5 | 31.6 | 44.8 | 44.7 | 43.7 | 49.3 | |
| P6 | 30.4 | 44.2 | 45.3 | 42.8 | 50.7 |
Files
journal-<a>-<b>.jsonl.gz, one per pairing, with agent a as entrant 0 and agent b as entrant 2. Each file is gzipped
JSON Lines: a header line, then one game per line. Games 2k and 2k+1 are played on the same board with the seats
swapped. manifest.json lists each file's games, wins and sha256 (of the uncompressed file).
HexSet seats a two-player game at a four-seat table with two seats locked, so per-player lists have four entries: entrants 0 and 2 play, and entrants 1 and 3 are empty.
| field | meaning |
|---|---|
index |
the game's number |
board_index |
the board pair (index // 2) |
cell |
the pairing, as in the file name |
policies |
the agent at each entrant (null for the empty ones) |
winner |
the winning entrant (0 or 2) |
seat |
the winning seat |
seating |
the seat of each entrant |
turns |
the number of turns |
points, roads, settlements, cities |
each entrant's final victory points and pieces |
record |
the HexSet game record |
Loading
import gzip
import json
from hexset.record import from_json, replay
with gzip.open("journal-P1-P3.jsonl.gz", "rt") as f:
for line in f:
game = json.loads(line)
if game["kind"] != "game":
continue
final = replay(from_json(json.dumps(game["record"])))
assert final.won_by == game["seat"]
hexset.record.replay_to(record, n) gives the position after n actions.
Licence
- Downloads last month
- 72