eai-v2-internal/agent-runs

Agent runs: replays and images

The replay videos and images of the agent runs shown on https://embodied-agent-interface-v2-internal.github.io/ (the "Agent runs" section). For every trial:

• replay.mp4: the separate verifier's replay of the trajectory the agent handed in. It is H.264 with the index first, so it plays and seeks in a browser before the whole file arrives. replay.webp is its last frame.
• img/: the images the model was shown during the run.
• snap/: only where the agent handed in no trajectory, the newest images it saved.

The site's log pages load these files from here, and each page carries the text of its trial: the model's words, its tool calls and their output, and the token counts.

Layout

<benchmark>/<run>/<task>/<slot>/replay.mp4
                               /replay.webp
                               /img/<n>.webp | <n>.jpg
                               /snap/<n>.webp | <n>.jpg
manifest.json                  every file, its size, and the size of the original it was compressed from

• benchmark: the site's benchmark id (behavior-1k, robolab, robowits, robopaint, …).
• run: the agent configuration as <harness>-<model>-<effort>[-<tag>], e.g. codex-gpt6_luna-xhigh: the same id as on the site and in the lab's results folders. The harness version and the route are in the site's run details, not in the id. (Until 2026-09-28 this run's folders were named codex-0.157-gpt6luna-xhigh-cgpt.)
• slot: the mode (unlimited, limited), or <mode>-prev for a run that a rerun replaced.

Adding your benchmark

Each benchmark owns its own folder. You add yours with your own account in this organisation:

1. Register and collect your runs in the site's repository:
   • state/runs/<benchmark>.yml and data/agents/<run>.yml;
   • make runs.
2. Publish the text snapshot with make publish-runs, and open a pull request with data/published_runs/.
3. Upload your media:
   • log in with hf auth login (your own account);
   • run make upload-run-media BENCHMARK=<benchmark>. It exports your benchmark's media, compresses it and uploads it to <benchmark>/ here, touching no other folder.
4. Merge. Once the pull request is merged, the site's log pages load your media from https://huggingface.co/datasets/eai-v2-internal/agent-runs/resolve/main/.

Licence

These files are renders from each benchmark's simulator and assets (BEHAVIOR-1K, RoboLab, RoboWits, RoboPaint, …). They stay under the terms of the benchmark they come from and of the assets shown in them. We grant no licence beyond those terms. Please ask before reusing them outside research.