Track the original source
Clips cut from the same session can share people, backgrounds, lighting, and near-identical moments. Request identifiers that link derivatives to their original recording. A unique filename by itself does not establish an independent example.
Choose a separation rule
Decide which kind of generalization matters. You might separate sessions, speakers, locations, productions, or recording dates. The right grouping depends on the intended evaluation; write it into the brief before anyone creates a random clip split.
Review near duplicates
Ask how repeated edits, exports, and overlapping intervals are identified. Consider the same source published in different resolutions or aspect ratios. Preserve source relationships in a manifest so that later transformations do not erase your split decisions.
Keep evaluation decisions reproducible
Record the grouping rule, source identifiers, and exclusions alongside the dataset version. Maintain a stable holdout where appropriate and document changes. This makes it easier to interpret model results when new footage or annotations are added.
Review checklist
- Model task and content categories
- Desired duration and total volume
- Video and audio specifications
- Overlays, edits, and content exclusions
- Commercial use, sample criteria, and timing