openfront
openfront-human-games
OpenFront.io archived human games
285 real public multiplayer games of OpenFront.io
(10-116 humans per lobby, all maps, FFA + team modes), pulled from the
game's public archive API and replayed through the deterministic engine
(openfrontio/OpenFrontIO) to
regenerate full state trajectories. ~420k snapshots at 10-tick (1s) cadence.
Every game passed the engine's embedded state-hash verification during
replay (games that desynced were discarded), so snapshots are bit-exact… See the full description on the dataset page: https://huggingface.co/datasets/djmango/openfront-human-games.openfront-snapshots
OpenFront.io game-state snapshots
250 bot/nation-only games of OpenFront.io (25 games on
each of 10 maps), played headless on the real deterministic game engine
(openfrontio/OpenFrontIO) and
snapshotted every 10 ticks / 1s of game time (~375k snapshots total).
Generated by and documented at
github.com/djmango/openfront-ai —
see the repo for loaders, training code, and results.
Format v3: every snapshot captures the complete game state — the tile
grid plus all entities (players… See the full description on the dataset page: https://huggingface.co/datasets/djmango/openfront-snapshots.openfront-replays
OpenFront.io RL game replays
Sparse-turn GameRecord v0.0.2 JSON for every finished RL episode
(training, watch, showcase). Same format as the OpenFront client archive /
Node bridge (version, gitCommit, sparse turns with intents).
Each parquet row is one game. The record_json column is the full Node-
compatible GameRecord string (~100–200 KB for long games). When a watch /
showcase episode recorded policy decisions, thinking_json holds a compact
top-3 action trace (typically a… See the full description on the dataset page: https://huggingface.co/datasets/djmango/openfront-replays.