faience-games

Faïence: human-vs-net Azul games

Every game played on Faïence, a free browser implementation of the rules of Azul (Michael Kiesling) against a neural net trained by self-play, unless the player switched sharing off. This dataset is the training pile the playing page tells its players about, and it is public precisely so that a player can read everything the project collects. Records are anonymous by construction: moves, deals, which net played, and the score. No names, no accounts, no IPs, no user agents.

Layout

games/YYYY-MM-DD/<timestamp>-<n>.jsonl, one file per ingest batch, one JSON object per line. Nothing is ever rewritten; new batches only add files.

Record format (faience-game/1)

Each line is a canonical record rebuilt by the collector (RemiFabre/faience-ingest), which replayed the game in the real engine and kept it only if the recorded deals, final scores and round count reproduce exactly. Fields:

• received_at (server clock, ISO) and created_at (client clock, may be null)
• seed: the game's RNG seed (mulberry32, the page's own RNG)
• human_seat, human_first: which of the two seats the human held
• net: {run, checkpoint, elo, params, backend} of the opponent
• think_time_s: the AI's search budget per move (0 = policy head only)
• moves: [{ply, player, action, sims?, value?}], action encoded as source*30 + color*6 + dest (identical in the JS and Python engines); sims is the positions the net searched for its move on the visitor's machine, value its root value on a [-1, 1] scale
• deals: per round, the five factories plus bag and lid counts, so a record replays independently of any RNG port
• final: {finished, scores, outcome, rounds, exhausted}; finished: false marks an abandoned game (position data with no outcome; train the value head on these with care, or not at all)

Caveats

• The collector deduplicates retried submissions by content, but a restart can rarely let a duplicate through: deduplicate by (seed, human_seat, moves, final.scores, final.finished) when it matters.
• An abandoned game that was later resumed and finished in the same tab can appear twice: once finished: false, once finished: true with the same seed and a longer move list. Prefer the finished one.
• Play strength varies wildly: these are self-selected browser visitors, from first-time players to strong club players.

Provenance and license

Collected by the Faïence ingest Space from the Faïence playing page; code and methodology in RemiFabre/ludometer. The records are dedicated to the public domain (CC0). Azul is a game by Michael Kiesling; this fan research project is not affiliated with or endorsed by its publishers.