adds simple first views on model callbacks

[jhoetter]

No description provided.

[jhoetter]

Run this together with:

• Proxy PR
• Gateway PR
• Submodule PR

You can add a model callback using the Python SDK in branch external-information-sources (PR) as follows:

from refinery import Client

user_name = "your-mail" # this is the email you log in with
password = "your-pw"
project_id = "your-clickbait-sample-project-id" # can be found in the URL of the web application

client = Client(user_name, password, project_id, uri="http://localhost:4455")

from sklearn.linear_model import LogisticRegression
clf = LogisticRegression()

from refinery.adapter.sklearn import build_classification_dataset
data = build_classification_dataset(client, "headline", "__clickbait", "distilbert-base-uncased", num_train=50)
clf.fit(data["train"]["inputs"], data["train"]["labels"])

from refinery.callbacks.sklearn import SklearnCallback
callback = SklearnCallback(
    client, 
    clf,
    "clickbait", 
)

callback.run(data["train"]["inputs"], data["train"]["index"])
callback.run(data["test"]["inputs"], data["test"]["index"])

[JWittmeyer]

Version overview doesn't seem to work on your branch

• resolved

[JWittmeyer]

websocket integration seems to be missing for stats - weren't refreshed on the overview page until I refreshed manually)
websocket integration seems to be missing for overview (newly created model callback)

• resolved

[JWittmeyer]

no option to change the comment or the name of the information source

• resolved
• ignored

[JWittmeyer]

lookup lists buttons have differing font weight

(actions might not be as straightforward since its created with the kern dropdown - making all bald would do the trick imo (also fits with other buttons on heuristic page)

• resolved

[JWittmeyer]

Maybe a UX issue we don't have to solve now.

For me the usage/integration feels a bit weird.
The model callback is somewhat hidden behind the heuristic page, the model callback block in the databrowser clearly depicts it as something else than a heuristic.

Also for clarity/differentiation, it might be a good idea to have a model page with callbacks & weak supervision combined (since weak supervision is basically our model and we want to compare it to a production model). Would also help to differentiate between the gravity of the results and give us the nice stat overview for the weak supervision :)

Another thing to help me identify it for what it is could be to add a hint in the databrowser or a graying it out if no model callback is present in the project.

• discussed
• issue created
• resolved

[JWittmeyer]

new weak supervision & model callback labels aren't vertically aligned

classes inline-flex items-center should do the trick. also padding right seems a bit low with pr-0.5 (tested with pr removed & px-1 instead of px-2)

• resolved