Paste a table of views — user_id, video_id, watch_time and optionally hour_of_day — and score them together with the same model. The bundled sample has twenty rows, two of them wrong on purpose.
Input
The service caps a batch at 100 rows and answers 400 above that; this page truncates the same way.
The one idea this station exists for: a batch is a collection of independent requests. Row 5 has a user that does not exist; row 12 has a video that does not exist. The other eighteen are scored normally and the two failures say why, in place, instead of the whole batch coming back as an error. The only thing that fails a whole batch is the model itself failing — because blaming that on a row would be a lie.