Ideally, a data collection should be large enough to build a working model.
Let’s use an example of creating a data collection of a certain type of pills (Ibuprofen) for our example.
If you want to build an ideal model, you need to start the first data set with 20 bottles of Ibuprofen from each manufacturer (ideally from different production lots), where you scan 5 pills from each bottle, and scan each pill three separate times.

3x5x20 = 300 total scans (for each manufacturer)
The result of the 300 scans is a baseline for establishing one type of pill from one manufacturer.
Next, you’ll do the same process for another pill from another manufacturer. Again, 20 bottles, where you scan 5 pills from each bottle, and scan each pill three separate times.

3x5x20 x 2 =600 total scans

From here, you’ll have two baselines from two manufacturers and can create your first model.