restrict model output to a defined character set?

[wsyxbcl]

Is there a way to customize the "dictionary" that the model uses, restricting the output characters to a specified subset? I understand that training a customized model might be the only choice, but I'm curious if there's any workaround.

[robertknight]

There isn't a built-in option to customize the output character set at present. If you just want to filter out certain characters you can post-process the outputs of recognize_text. If you want to force the model to choose between only an allowed set of characters (for example, always predict numbers), you'd probably want to fine-tune an existing model by starting from an existing checkpoint and replacing the final layer. I'm afraid there isn't a guide specifically for fine-tuning yet, but it would be similar to the training guide.

There is a related issue for supporting other languages/scripts (which could include subsets of the default alphabet) here.

[wsyxbcl]

Thanks for the detailed reply, will try fine-tuning the model and let you know the result.