Bug Report: Mask token mismatch with the model on hosted inference API of Model Hub

Beginners
1

In my model card, I used to be able to run the hosted inference successfully, but recently it prompted an error: "<mask>" must be present in your input.

My model uses RoBERTa MLM and BERT Tokenizer. So the mask token is actually “[MASK]”. I have already set it in tokenizer_confg.json but the inference API still mismatches with that.

In the past it is OK but recently it turns to prompt an error. Seems like the front-end start to double-check the mask token. How can I set the mask token in an appropriate way? Is it documented to set mask token in inference API?

Thanks!

To reproduce

Steps to reproduce the behavior:

  1. Go to ethanyt/guwenbert-base · Hugging Face
  2. Run an example with “[MASK]”

Expected behavior

In the past it was OK. See snapshot in guwenbert/README_EN.md at main · Ethan-yt/guwenbert · GitHub

2

Should be resolved in Mask token mismatch with the model on hosted inference API of Model Hub · Issue #11884 · huggingface/transformers · GitHub, but if possible do not open duplicate issues/forum posts. Thanks!

2 Likes