Starting from published models.
WHAT THEY PROVIDE
Pre-trained weights Standardised loading interfaces Datasets Model documentation
WHY THEY CHANGED PRACTICE
Starting from scratch became unnecessary for most tasks.
WHAT TO CHECK BEFORE USING A MODEL
The licence, and whether it permits your use What data it was trained on Its documented limitations Its size and inference cost Whether it is maintained
WHY THE LICENCE SPECIFICALLY
Some permit research only, and commercial use would breach them.
WHAT TO BE CAREFUL WITH
Models loading code on initialisation Models from unverified publishers Weights whose provenance is unstated
WHY THAT FIRST POINT MATTERS
Loading a model file can execute arbitrary code.
WHAT TO PREFER
Formats that do not execute code, and publishers you can identify.
WHAT TO PIN
The exact model version.
WHY
Published models are updated, and behaviour changes.
WHAT TO EVALUATE
Performance on your own data, before adopting.
WHY
Reported benchmark performance frequently does not transfer.
WHAT TO RECORD
Which model, which version, and when it was adopted.
