Stereotype Judge

Type a sentence and the classifier labels it a gender stereotype, anti-stereotype, neutral, or not gender-related. Click Explain to see which words drove the decision.

One sentence per line (up to 20). Ctrl/⌘ + Enter to classify.
How it works
  1. Gender filter. A Snowflake-L sentence embedding and a logistic regression give P(gender-related). Below 0.5 the sentence is labeled "not gender-related" and the judge's scores are shown greyed out.
  2. Judge. Qwen2.5-14B (4-bit) with a LoRA adapter gives P(anti-stereotype), P(stereotype) and P(neutral); the largest is the label.
  3. Explain. Word-level Shapley values for the judge's log-odds of its label: each word is hidden ("...") in many combinations and the change in the score is averaged. Orange words push toward the label, blue words push away. Hidden-word sentences are unusual text, so treat this as a rough guide.

The training labels are relative: "stereotype" means more stereotypical than its paired sentence. Comparing a sentence with its gender swap is often more informative than a single score.