djmango's picture
Upload README.md with huggingface_hub
25e3c53 verified
|
Raw
History Blame
2.03 kB
metadata
license: mit
pretty_name: OpenFront.io archived human games
tags:
  - reinforcement-learning
  - imitation-learning
  - 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 reconstructions of what the players actually saw. Player identities are stripped by the API; usernames remain.

Generated by and documented at github.com/djmango/openfront-ai.

Contents

  • maps/<map>.tar — replayed snapshots, one game directory per game, same format v3 as openfront-snapshots: terrain.bin, states/t<tick>.bin.gz (tile ownership grid), states/t<tick>.json.gz (players, diplomacy, units, attacks), meta.json (+ gitCommit, hashesChecked, numHumans).
  • records/<gitCommit>/<gameID>.json.gz — the raw archived game records: full per-turn human intent logs (every attack, boat, build, nuke, alliance, betrayal, donation, emote). Replay them with datagen/replay.ts from the repo (engine checked out at the matching commit), or use them directly for behavior cloning.

Why

Bot self-play data (openfront-snapshots) covers the state space thinly: built-in nations rarely ally, betray, nuke, or invade by sea. Human games supply realistic territory shapes and the full diplomatic/military action distribution — both better observation coverage for encoders and (state, human action) pairs for imitation pretraining.