Ian Tenney

We don't have a generic wrapper for this, but you should be able to implement as a custom frontend module; see https://github.com/PAIR-code/lit/blob/main/documentation/development.md#custom-client--modules and https://github.com/PAIR-code/lit/blob/main/documentation/client.md for more details.

So for an NER model, you probably want to have a `SpanLabels` or `SequenceTags` field for the model predictions. So your output spec would look something like: ``` def output_spec(self)...

Unfortunately, we don't currently have a type for the full output probabilities of that form. As an approximation though, you could use `TokenTopKPreds` for per-token probabilities, which would be a...

I'm not familiar with this debugger, but it looks like `list_result` in the second image is a single-element list, which should be correct.

FWIW, I found a reasonable workaround was to just add two legends, one for each column. ``` legend_items = list(zip(names, renderers)) legend_kw = dict(spacing=15, label_text_font_size="13pt", orientation='vertical') legend = bokeh.models.Legend(items=legend_items[:2], location=(10,0),...

I have not seen this before, possible that it's ingesting an empty batch somehow, thus `num_tokens == []`?

I think it would be possible. Image input should work fine, but you may need to do some custom UI work (https://github.com/PAIR-code/lit/wiki/frontend_development.md#custom-client--modules) to show the bounding boxes. If you wanted...

You can use tensor data for `Embeddings`, `TokenEmbeddings`, or similar types - you will just want to be sure to convert it to numpy arrays as LIT expects plain-old-data that...