Nico-robot/duckjam-dw2-vault

The Duck Jam Vault

Every submission entered in Duck Jam and every run the Arena scored, as two Parquet tables and one JSON board per round. A publisher job rewrites this repository once a day, so its git history is the history of every leaderboard.

This repository is a view. The rows themselves are written, once and never edited, into public Hugging Face Buckets by the components that make them: hf://buckets/Nico-robot/duckjam-submissions, hf://buckets/Nico-robot/duckjam-runs, hf://buckets/Nico-robot/duckjam-erasures. Anyone who distrusts the tables below can list those buckets and fold the same numbers out of them — that is the point of publishing both.

What is in here

path	what it holds
submissions/<yyyy-mm>.parquet	one row per submission created that month
runs/<yyyy-mm>.parquet	one row per evaluation that finished that month
leaderboards/<challenge_id>/<major>.json	the board of one round, folded at the version its runs have reached
rounds/<challenge_id>/<version>/	the released round: the package the Arena pinned, and the composed scene of every evaluation seed

Both tables are sorted by created_at, then by id, and every row carries the exact bucket document it came from in its row column — so a reader who wants the seed-by-seed scores, the trace keys or the gate's checks reads json.loads(row) and has the whole thing.

Reading it

hf datasets sql "
  SELECT owner_name, aggregate FROM 'hf://datasets/Nico-robot/duckjam-vault/runs/*.parquet' r
  JOIN 'hf://datasets/Nico-robot/duckjam-vault/submissions/*.parquet' s USING (submission_id)
  WHERE r.status = 'scored' ORDER BY aggregate DESC LIMIT 10"

import duckdb
duckdb.sql("SELECT * FROM 'hf://datasets/Nico-robot/duckjam-vault/runs/*.parquet' WHERE status = 'scored'")

A leaderboard is a fold over these rows and is never stored as state: the newest scored run under the round's current version counts for a submission, the best aggregate counts for a player, and ties go to the earliest submission. The published leaderboards/ files are that fold, run by the publisher; arena/vault/board.py is the same function in eighty lines, and the Dome and duckjam board call it too.

Erasure

An erasure row hides the runs it names from every board and orders their traces, logs, artifact and media deleted. The publisher drops those runs from the Parquet on its next pass and a daily job removes their files, so a run that has been erased is absent here while the record that it once existed — its row, and the erasure row — stays in the buckets.

Provenance

Scores are comparable across a round because a run row names everything that produced it: the challenge version and the sha256 of its package, the runner image, and the servo model the actuator used. A run scored under an older version of a round is kept, superseded, and no longer counted — supersedes and reason say which run replaced it and why.