sentence_starts is empty list and the first element of sentence_starts is not 0.

[changzhisun]

I found that “sentence_starts” in the qed-train.jsonlines is empty for some document. Is this a annotation error ？
And In some document, the first element of sentence_starts is not 0. Is the previous text not required？

[calberti]

Thanks, we should fix this. The sentence boundaries were generated by an older model and are noisy, but for consistency we should always have 0 as a sentence start.