Distilled / faster coreference resolution

[znadrich-qf]

In comparison to fastcoref the en_coreference_web_trf coreference resolution model available in spacy-experimental is significantly slower. There was a previous discussion about potentially increasing the model inference speed by using a distilled transformer. Before undertaking this effort myself, I needed to get ahold of the OntoNotes dataset which needs to be licensed. There is another discussion stating that spaCy has a special licensing agreement allowing you to release models trained with this dataset under the MIT license

Given all of this, it seems rather difficult for one to train a distilled coreference resolution model without the dataset or the specialized licensing agreement. Would spaCy ever officially release a distilled coreference resolution model?

[znadrich-qf]

Any thoughts on this?