VLABench runs¶

A run is one agent configuration, harness + model + settings; each task in it runs once per mode. Pick a run above; the table lists every task of it, one row per mode, in the default order (finished first, then running, easier tasks first). A pill links to that trial's log page. Filter the table with the chips above it.

successfailederror / stalledrunninggradingqueuednot run

Bars count trials, one per task and mode, by state. Progress is the partial credit beside success, in [0, 1]: BEHAVIOR's 2026 challenge q_score, elsewhere the grader's final_reward. A success always counts as full progress, 1.0, whatever that key says (RoboLab's final_reward is 0 for a task without scored subtasks); mean progress averages it over the graded trials. Agent time adds up finished trials; running ones are shown apart. A benchmark can also declare grader metrics, continuous scores its grader writes (IoU, F1, a distance): they get a column each, a sort, and a run figure, the mean or the median of the graded trials.

Tokens are the model gateway's count of every request (input includes the cached part, output the reasoning part). ≈ $ is those tokens at the list prices in data/prices.yml, an estimate and never a bill; billed is what a provider charged, where it says. The two are never added together.

A benchmark's owner can exclude tasks from the final benchmark (excluded: in state/tasks/<benchmark>.yml). Their trials stay, last in a run's table under Excluded from the benchmark; the numbers count the benchmark's own tasks, and the Count switch adds the excluded ones.

Runs are registered per benchmark: how to register a run.