Update datasets

[gabegma]

Resolve #267

Description:

Checklist:

You should check all boxes before the PR is ready. If a box does not apply, check it to acknowledge it.

• ISSUE NUMBER. You linked the issue number (Ex: Resolve #XXX).
• PRE-COMMIT. You ran pre-commit on all commits, or else, you
  ran pre-commit run --all-files at the end.
• USER CHANGES. The changes are added to CHANGELOG.md and the documentation, if they impact
  our users.
• DEV CHANGES.
  • Update the documentation if this PR changes how to develop/launch on the app.
  • Update the README files and our wiki for any big design decisions, if relevant.
  • Add unit tests, docstrings, typing and comments for complex sections.

[gabegma]

@Dref360 do you think my fixes for updating datasets make sense?

[Dref360]

looks good. We never run anything on the DatasetDict themselves so there is nothing to worry about if they don't have the same features.

[JosephMarinier]

Is it possible to instantiate only one ClassLabel object and pass it to both dataset managers?

Suggested change

	eval_dm._base_dataset_split.features["label"] = ClassLabel(
	num_classes=3, names=existing_classes + ["NO_INTENT"]
	)
	train_dm._base_dataset_split.features["label"] = ClassLabel(
	num_classes=3, names=existing_classes + ["NO_INTENT"]
	)
	eval_dm._base_dataset_split.features["label"] = train_dm._base_dataset_split.features[
	"label"
	] = ClassLabel(num_classes=3, names=existing_classes + ["NO_INTENT"])

Or with a temporary variable:

Suggested change

	eval_dm._base_dataset_split.features["label"] = ClassLabel(
	num_classes=3, names=existing_classes + ["NO_INTENT"]
	)
	train_dm._base_dataset_split.features["label"] = ClassLabel(
	num_classes=3, names=existing_classes + ["NO_INTENT"]
	)
	class_label = ClassLabel(num_classes=3, names=existing_classes + ["NO_INTENT"])
	eval_dm._base_dataset_split.features["label"] = class_label
	train_dm._base_dataset_split.features["label"] = class_label

Or if we move the creation of dms before that, we can loop on it:

Suggested change

	eval_dm._base_dataset_split.features["label"] = ClassLabel(
	num_classes=3, names=existing_classes + ["NO_INTENT"]
	)
	train_dm._base_dataset_split.features["label"] = ClassLabel(
	num_classes=3, names=existing_classes + ["NO_INTENT"]
	)
	class_label = ClassLabel(num_classes=3, names=existing_classes + ["NO_INTENT"])
	for dm in dms.values():
	dm._base_dataset_split.features["label"] = class_label

Here is a complete diff of the last idea, in case it was not clear:

     # Adding a rejection class
     eval_dm: DatasetSplitManager = mod.get_dataset_split_manager(DatasetSplitName.eval)
     train_dm: DatasetSplitManager = mod.get_dataset_split_manager(DatasetSplitName.train)
-    existing_classes = eval_dm.get_class_names(labels_only=True)
-    eval_dm._base_dataset_split.features["label"] = ClassLabel(
-        num_classes=3, names=existing_classes + ["NO_INTENT"]
-    )
-    train_dm._base_dataset_split.features["label"] = ClassLabel(
-        num_classes=3, names=existing_classes + ["NO_INTENT"]
-    )
-    eval_dm._base_dataset_split = eval_dm._base_dataset_split.map(
-        lambda u, i: {"label": 2 if i % 10 == 0 else u["label"]}, with_indices=True
-    )
     dms = {
         DatasetSplitName.eval: eval_dm,
         DatasetSplitName.train: train_dm,
     }
+    existing_classes = eval_dm.get_class_names(labels_only=True)
+    class_label = ClassLabel(num_classes=3, names=existing_classes + ["NO_INTENT"])
+    for dm in dms.values():
+        dm._base_dataset_split.features["label"] = class_label
+    eval_dm._base_dataset_split = eval_dm._base_dataset_split.map(
+        lambda u, i: {"label": 2 if i % 10 == 0 else u["label"]}, with_indices=True
+    )

[gabegma]

Thank you for the recommendation! I further cleaned it up, because I think it was hard to read, given that sometimes we would edit the values in the Dict, and sometimes, directly eval_dm. LMK what you think.

[JosephMarinier]

Haha! I had done the exact same change locally while reviewing, but I thought I was asking too much. That's perfect! 👍

[JosephMarinier]

That's a relief! Thank you! I have some minor comments, but that's good to go!