Three numbers an operator checks in the first five seconds after a deployment — the artefact's type, its input width, its version — then a chart of what the model actually listens to.
Status: loaded · metrics {"auc": 0.563204, "accuracy": 0.71895}
[ "watch_time", "hour_of_day", "day_of_week", "is_weekend", "month", "duration_seconds", "days_since_upload", "view_count", "like_count", "comment_count" ]
Pipeline means the data preprocessing is packed inside the same file as the model, so the two cannot drift apart. 54 is the number of raw inputs — if the last release said 54 and this one says 32, a data join broke and nothing else would have told you. The version is whether this is the file you meant to ship. A service can pass every health check with last month's model loaded; this page is how you would know.| feature | importance | coefficient |
|---|---|---|
| num__watch_time | 0.2307 | +0.2307 |
| cat__is_premium_true | 0.0885 | -0.0885 |
| cat__has_music_true | 0.0885 | -0.0885 |
| cat__is_vertical_false | 0.0849 | -0.0849 |
| cat__has_text_overlay_true | 0.0843 | -0.0843 |
| cat__language_en | 0.0834 | -0.0834 |
| cat__is_monetized_true | 0.0810 | -0.0810 |
| cat__has_effects_false | 0.0806 | -0.0806 |
| cat__video_quality_720p | 0.0805 | -0.0805 |
| cat__has_voiceover_true | 0.0800 | -0.0800 |
| cat__preferred_category_Beauty | 0.0785 | -0.0785 |
| cat__has_voiceover_false | 0.0744 | -0.0744 |
| cat__video_quality_1080p | 0.0739 | -0.0739 |
| cat__has_effects_true | 0.0739 | -0.0739 |
| cat__is_monetized_false | 0.0735 | -0.0735 |
| cat__top_country_US | 0.0728 | -0.0728 |
| cat__has_text_overlay_false | 0.0702 | -0.0702 |
| cat__is_vertical_true | 0.0695 | -0.0695 |
| cat__preferred_category_Kids | 0.0666 | -0.0666 |
| cat__has_music_false | 0.0660 | -0.0660 |
num__watch_time at 0.231 — is 2.6× the next one, and the rest are a crowd of one-hot video flags at a height that is visibly noise. This model was deliberately trained on 54 inputs so this page has something to show; the chart is the visual version of the earlier finding that only watch time predicts engagement here. The coefficient's sign is the direction: negative means that input lowers the odds, given the others. Names like num__ and cat__ are the preprocessor's — proof it is inside the artefact.