Alan Akbik
Alan Akbik
Following the comment in #2888, experiment with alternative heuristic that distinguishes between embeddings and all other parts of the model.
`TransformerWordEmbeddings` with `distilbert-base-uncased` cannot embed an empty token if at the end of a sentence: ```python # sentence with two empty tokens, one at end sentence = Sentence(["hello", " ",...
Time to do another release. This issue tracks the final changes and progress of releasing.
Hello all, we now have another **position for a PhD candidate** available! It is a full-time and fully paid research associate position in my group, intended for persons with a...
This issue tracks the process of bugfix release 0.11.3 - mainly fixing the problems caused by the missing `sacremoses` dependency in the new `transformers` version.
Make an incremental release (version 0.11.2) - the release covers the bugfixes and updates of the last weeks. Functionality-wise, it also adds to `SequenceTagger` the option of predicting tags at...
### Describe the bug The current implementation of `TransformerEmbeddings` makes the assumption that between two tokens there is always a whitespace. As a consequence, the following two sentences get the...
When using Flair with Torch 2.0, our "oldest" models can no longer be loaded. This is because they serialized LSTM properties that no longer exist in Torch 2.0. This issue...
This issue tracks the progress of releasing the next version of Flair, i.e. Flair 0.13. Planned release is **16.10.23**. One major new documentation feature are **publicly available API docs**, generated...
This issue tracks the progress of adding support for the Ukrainian language from [lang-uk](https://lang.org.ua/en/) to Flair. We would like to add: - [x] Ukrainian Flair embeddings trained by @dchaplinsky and...