MLL-Lab/embodied-exam-runs

Embodied First Exam: agent runs

The agent runs behind the run viewer of the Embodied First Exam, a benchmark for code agents in robotics: every step, tool call and picture an agent saw while it sat a task, the verifier's replay of its final answer, and the agent's full trajectory for download.

Layout

index.json                              runs per suite, solved per suite, runs per setting
index/<suite>.json                      one row per run of the suite
runs/<h2>/<slug>/run.json               one run: every step, tool call and output
runs/<h2>/<slug>/replay.mp4             the verifier's replay of the final answer (when there is one)
runs/<h2>/<slug>/poster.jpg             the replay's last frame
runs/<h2>/<slug>/steps.bin              the pictures the agent saw, concatenated; run.json gives each as [offset, length]
runs/<h2>/<slug>/trajectory.jsonl.gz    the agent's full trajectory, one event per line

h2 is the first two hex characters of sha1(slug). A slug is <task dir>--<last 7 characters of the trial id>. index.json also names, under linked, a few earlier runs that the website links directly; they are kept here but are not listed in the shards.

Contributing

The runs come from the exam's own evaluations. Please do not open pull requests on this repository.

Licence

The trajectories, pictures and replays come from agent runs on the exam's tasks, made by the exam's maintainers. The scenes and assets keep the licences of their upstream benchmarks; see each suite's page on the exam's website.