Submitting a method
What your method may see
The film's footage, and nothing else. No rubric, no reference stills, no metadata, no chains. Multimodal methods additionally get one fixed sentence, identical for every film:
Pick a hero frame from this movie. A hero frame is a single still image taken from within a movie that would work as its cover, thumbnail, or promotional entry point.
Available as heroframe.QUERY. Use it verbatim.
Frames must come from the shared 1 fps track. Selecting off a denser grid of your own turns the comparison into a decoding-density comparison.
Output format
One JSON object per line, one line per film:
{"movie_id": "0001_American_Beauty", "frames": [{"rank": 1, "frame_idx": 2477}, {"rank": 2, "frame_idx": 1893}]}
| Field | Required | Meaning |
|---|---|---|
movie_id |
yes | as in the benchmark |
frames[].frame_idx |
yes | index into that film's track |
frames[].rank |
yes | your own ordering, 1 = best. S@k measures how well you rank your own output |
frames[].score |
no | your internal score, carried through for analysis |
At most 5 frames per film. You choose how many. Scoring averages across your frames rather than taking the best, so a fifth mediocre frame can lower your score — submit the ones you would actually stand behind.
Before you submit
Verify your frame track. python scripts/02_verify_track.py --track ./track
should report all films verified. A pool that differs from ours produces a
number that is not comparable, and nothing downstream will warn you.
Score random as a control. On the full corpus it should land at 0.501. If your pipeline gives something materially different, the problem is in the pipeline rather than in your method.
Do not modify the prompt. Scoring refuses to run if it has drifted. If you edited it deliberately to study the metric, say so and report the result as your own measurement rather than a benchmark score.
Name the internal model. If your method calls an LLM or VLM internally, put the model in the method name and report the published configuration. This is not bureaucracy: substituting CueKFS's internal model moved it across the random baseline, because a stage silently degraded into its fallback path without raising anything.
Opening the pull request
Add a row to LEADERBOARD.md and attach the result JSON from
scripts/04_evaluate.py. Include:
- Method name, with the internal model if there is one.
Sand the 95% CI, from the result JSON.- Reader used. Should be
Qwen3.6-27B; if not, say so — frame-level scores are not comparable across readers. - Number of films and units. Full corpus is 204 films; a subset is fine, but say which.
- Link to your code, or the selections file.
- Verification output from step 2.
Reporting a partial run
Scoring a subset is a legitimate thing to do — the full corpus is about 20,000 reader calls per method. Report which films, and compare against the same subset rather than against the full-corpus numbers:
python scripts/04_evaluate.py --selections mine.jsonl --track ./track \
--name mine --movies 0001_American_Beauty 0003_CASABLANCA 0008_Fargo
Per-film scores for every official row are in eval/baselines.jsonl, so you can
reconstruct any subset's reference numbers without re-running them.
What a strong result looks like
The gap to close is from 0.586 to 0.722 — 0.136 of S, or 0.818 of an insertion position. Two things about that gap, from the analysis:
It is not photographic quality. Brightness, contrast, sharpness, colourfulness and edge density all correlate at |ρ| ≤ 0.09 with a frame's score, and a selector optimising exposure and focus scores below random.
It is only partly faces. Face size is the one strongly correlated attribute (ρ = 0.485), but it is a threshold rather than a dial and it tops out at 0.557. Human publicity stills score 0.165 higher with faces one eighth the size.