The inferenced text has a leading space because of the vocab

[Zardinality]

Debugged to the first token emerged:

And confirmed by using python model directly:

Clearly the tokenizer model count ▁Once instead of Once as a token. Note that sentencepiece use ▁ as space.

[karpathy]

Yes, the vocab itself has this space. But somehow when you ask sentencepiece to decode a token sequence, it doesn't print this leading space.

[Zardinality]

I am not familiar with the tokenizer common practices, but seems there it has a patch for such scenario in decoding: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_processor.cc#L788 . Might have a good reason to do so.

[karpathy]

This was fixed while back