Remove the toxicity demos from the LIT examples.

PiperOrigin-RevId: 640521996

lit_nlp/examples/datasets/classification.py

@@ -187,35 +187,3 @@ def spec(self) -> lit_types.Spec:
         "label": lit_types.CategoryLabel(vocab=self.LABELS),
     }
 
-
-class ToxicityData(lit_dataset.Dataset):
-  """Jigsaw toxicity dataset; see https://www.tensorflow.org/datasets/catalog/wikipedia_toxicity_subtypes."""
-
-  LABELS = ["non-toxic", "toxic"]
-
-  def __init__(self, split="test", max_seq_len=500):
-    """Dataset constructor, loads the data into memory."""
-    raw_examples = load_tfds("wikipedia_toxicity_subtypes", split=split)
-    self._examples = []  # populate this with data records
-    for record in raw_examples:
-      self._examples.append({
-          "sentence": record["text"].decode("utf-8"),
-          "label": self.LABELS[int(record["toxicity"])],
-          "identity_attack": bool(int(record["identity_attack"])),
-          "insult": bool(int(record["insult"])),
-          "obscene": bool(int(record["obscene"])),
-          "severe_toxicity": bool(int(record["severe_toxicity"])),
-          "threat": bool(int(record["threat"]))
-      })
-
-  def spec(self) -> lit_types.Spec:
-    """Dataset spec, which should match the model"s input_spec()."""
-    return {
-        "sentence": lit_types.TextSegment(),
-        "label": lit_types.CategoryLabel(vocab=self.LABELS),
-        "identity_attack": lit_types.Boolean(),
-        "insult": lit_types.Boolean(),
-        "obscene": lit_types.Boolean(),
-        "severe_toxicity": lit_types.Boolean(),
-        "threat": lit_types.Boolean()
-    }

lit_nlp/examples/models/glue_models.py

@@ -648,25 +648,3 @@ def input_spec(self):
     ret = super().input_spec()
     ret[self.config.label_name] = lit_types.Scalar(min_val=0, max_val=5)
     return ret
-
-
-class ToxicityModel(GlueModel):
-  """Classification model on Jigsaw Toxicity Dataset."""
-
-  def __init__(self, *args, **kw):
-    super().__init__(
-        *args,
-        text_a_name="sentence",
-        text_b_name=None,
-        labels=["non-toxic", "toxic"],
-        null_label_idx=0,
-        **kw)
-
-  def output_spec(self) -> Spec:
-    ret = super().output_spec()
-    ret["probas"] = lit_types.MulticlassPreds(
-        parent=self.config.label_name,
-        vocab=self.config.labels,
-        null_idx=self.config.null_label_idx,
-        threshold=0.3)
-    return ret

lit_nlp/examples/tools/glue_trainer.py

@@ -25,7 +25,6 @@
 from absl import flags
 from absl import logging
 
-from lit_nlp.examples.datasets import classification
 from lit_nlp.examples.datasets import glue
 from lit_nlp.examples.models import glue_models
 from lit_nlp.lib import serialize
@@ -130,10 +129,6 @@ def main(argv: Sequence[str]) -> None:
     train_data = glue.STSBData("train")
     val_data = glue.STSBData("validation")
     model = glue_models.STSBModel(_ENCODER_NAME.value)
-  elif _TASK.value == "toxicity":
-    train_data = classification.ToxicityData("train")
-    val_data = classification.ToxicityData("test")
-    model = glue_models.ToxicityModel(_ENCODER_NAME.value)
   else:
     raise ValueError(f"Unrecognized task name: '{_TASK.value:s}'")
 

website/sphinx_src/components.md

@@ -116,9 +116,7 @@ implemented with the `MulticlassPreds` and `CategoryLabel` types.
 *   A negative class can be designated using the `null_idx` attribute of
     `MulticlassPreds` (most commonly, `null_idx=0`), and metrics such as
     precision, recall, F1 will be computed for the remaining classes. AUC and
-    AUCPR will be computed for binary classification tasks. For an example, see
-    the
-    [comment toxicity model](https://github.com/PAIR-code/lit/blob/main/lit_nlp/examples/models/glue_models.py?l=518&rcl=386779180).
+    AUCPR will be computed for binary classification tasks.
 *   If `null_idx` is set and there is only one other class, the other class
     (often, class `1`) is treated as a positive class, and the LIT UI can be
     used to change the classification threshold. If `null_idx` is set and there