Round

—

State

—

Seats filled

—

Tasks

—

Head

—

Winner bar

—

Output cap, tokens

—

Next round

—

Public math benchmarks, scores in percent. One row for every king, newest first; a king is the head that one applied round made. Under the kings: the reference models, the largest first, and the genesis head, which is the model that v2 started from. A reference model is in view when it leads for its size: in the average or in one benchmark, no reference model of the same size or smaller has a higher score. The link at the end shows all. A reference model is a model of another lab; the benchmark service of v1 made its scores, and they are in the v1 history. A king with no score yet still has its row: its name says if the run is queued, is running, failed, or is not in the benchmark file yet. The cards show each score over time. A star marks the highest score of a column. † marks a partial run: some samples were lost, and each lost sample counts as wrong. For information only: no king and no weight depends on this table.

One position per closed round, the oldest at the left, for the rounds that mining.json lists. z is the gain over the head M in standard errors: the z of the applied winner, or the best z of a round with no winner. The pass rates are on the tasks of that round, so two rounds have different tasks. Duration is the time from round open to round close.

A king is a round winner that was applied to the head. The table holds the newest win of each of the last distinct winner UIDs. Each one is paid an equal share while its hotkey is still registered on that UID. Only a winner's batch and LoRA are public.

Seats in the round that runs now come first, then the waiting line in its order. A seat with no valid file is skipped, is not charged, and keeps its place. Queued files stay private.

One row per batch per round, newest first. A batch that did not win stays private: only its hash and its scores are shown.

Scores per environment. "M pass" is the pass rate of the head M with no adapter, on the same tasks and seeds as every LoRA of that round.

The head is a base checkpoint plus an ordered list of applied LoRAs. A fold merges the LoRAs into a new base.