A developer loads a text model whose tokenizer differs from the tokenizer used during training. Responses become nonsensical. What should be checked first?

1. Change only the web page background color.
2. Disable every log before further diagnosis.
3. Sort the checkpoint filenames alphabetically.
4. Load the tokenizer and special-token configuration compatible with the model checkpoint.