Datasets:

MLL-Lab
/

embodied-exam-data

The numbers behind the leaderboard, one row per counted attempt, and a link from every attempt to its full run. Snapshot of 5 November 2026; the newest counted attempt finished on 2 November 2026. Task set: the exam repository mll-lab-nu/embodied-first-exam at main ebd10b62b0.

What is counted

	tasks	tests (task x mode)	privileged	standard	realtime	suites
all	1,156	2,288	1,047	1,155	86	35
core	456	901	391	456	54

Core = docs/core_set.toml of that commit. Each core task has a default mode: 403 standard, 53 realtime.
One attempt per (agent, test): the latest finished valid attempt, success or failed. Infrastructure failures caught before grading are re-run and do not count; ablation and quick-check runs are left out. An attempt that an API or infrastructure error ended after grading keeps its verdict and is labelled in its run.
Success = verifier reward >= 1 within the token budget.
Missing counts as failed: a test with no counted attempt is a failure (59 of the 6,307 pairs).
Views: Core = each core task's default test (456). Privileged / Standard / Realtime = every core test of that mode (391 / 456 / 54). Rate = solved / tests of the view.

agent	Core /456	Privileged /391	Standard /456	Realtime /54	missing
Codex · GPT-6 Astra	150 (32.9%)	303 (77.5%)	156 (34.2%)	17 (31.5%)	3
Claude Code · Opus 5.5	107 (23.5%)	231 (59.1%)	114 (25.0%)	4 (7.4%)	1
Antigravity · Gemini 4 Argon	89 (19.5%)	164 (41.9%)	96 (21.1%)	4 (7.4%)	0
Antigravity · Gemini 3.8 Flash	27 (5.9%)	96 (24.6%)	27 (5.9%)	1 (1.9%)	0
Claude Code · Qwen3.8 Max	18 (3.9%)	37 (9.5%)	18 (3.9%)	1 (1.9%)	48
dsh · DeepSeek V4.1 Flash	12 (2.6%)	69 (17.6%)	12 (2.6%)	1 (1.9%)	0
Claude Code · GLM 5.3 Flash	11 (2.4%)	36 (9.2%)	11 (2.4%)	0 (0.0%)	7

"missing" counts tests over all three modes.

Files

file	content
`numbers.json`	the sizes above, the rule, and per agent and view: solved, tests, rate, missing
`leaderboard_core.csv`	901 core tests x 7 agents = 6,307 rows: verdict (`success` / `failed` / `missing`), `counts_as`, the attempt's `trial_id` and `hf_slug`
`trials.csv`	the 6,248 counted attempts: suite, task dir, setting, is_default, run, verdict, reward, finished_at (UTC), `in_common7` (the 843 core tests all seven agents have a result on), `hf_slug`, `has_video`, `hf_run_dir`, `viewer_url`, `video_url`
`tests.csv`	the 901 core tests: default or not, in the common set or not, and which runs lack a result
`check_coverage.py`	checks that every row of `trials.csv` has its run in MLL-Lab/embodied-exam-runs with the same outcome

run_id is <harness>-<model>-<effort>, e.g. codex-gpt6_astra-high. A task dir is <suite>-<task>-<mode>.

Opening a run

Every counted attempt has a run in the public dataset MLL-Lab/embodied-exam-runs:

slug = <task dir>--<last 7 characters of trial_id>, lower case (the hf_slug column);
viewer: https://www.embodied-first-exam.ai/run/?r=<slug>;
files: runs/<first 2 hex digits of sha1(slug)>/<slug>/: run.json (every step, tool call and output), trajectory.jsonl.gz (the agent's full trajectory), steps.bin (the images the agent saw), and replay.mp4 + poster.jpg when the attempt has a video (privileged attempts that submitted nothing have none).

huggingface-cli download MLL-Lab/embodied-exam-runs --repo-type dataset \
  --include "runs/e1/ambench-cabinet-pick-place-fahexa-i00-privileged--zvvkvgc/*" --local-dir efe-runs
python3 check_coverage.py trials.csv      # no token needed; HF may rate-limit anonymous calls for a few minutes

At MLL-Lab/embodied-exam-runs@dfb89b3 the check passes for all 6,248 counted attempts (5,839 with a video).

How this maps to the website

Leaderboard page: numbers.json per_agent (Core and the three mode views).
Task pages: each task's Results row = leaderboard_core.csv (S = success, F = failed or missing).
"Watch the run" links and the /run/ page: viewer_url, served from MLL-Lab/embodied-exam-runs.
A run whose last attempt an API or infrastructure error ended is shown with its verdict as counted and labelled ("Ended by an API error"); its run.json has ended_by.