Separate source records from training units
A source episode contains recorded observations and actions; a training window is a derived selection from those records. Ask how windows are sampled and whether they overlap. A collection with many windows can still contain relatively few original attempts. Keep source episode references so the buyer can distinguish additional sampling from genuinely additional task experience.
Define the prediction target
A model may predict one action or a sequence of future actions. Specify what each target step means, how timing is represented, and how the deployment controller will use it. Consecutive values should not be assumed to have a fixed time interval if source rates vary or steps were removed. Document any conversion to a regular sampling schedule.
Handle boundaries and missing steps
Windows near episode ends may be shortened, padded, or excluded. Describe that policy and provide masks or other indicators where needed. Prevent windows from silently crossing resets, interventions, or unrelated tasks. An apparently continuous action sequence can be misleading when the export joins segments that were not executed as one uninterrupted attempt.
Preserve evaluation independence
Overlapping windows from one episode should follow the intended source-level partition rule. Splitting derived windows randomly can place nearly the same behavior in training and evaluation. Keep sampling settings and dataset versions available so results can be reproduced. Include the desired unit and horizon in a sourcing brief only after clarifying which source signals and timing evidence are essential.
Review checklist
- Keep original episode references
- Define target sequence timing
- Document padding and boundary rules
- Group overlapping windows consistently