auto-debias

app.py

@@ -0,0 +1,42 @@
+import gradio as gr
+from transformers import pipeline
+
+import gc
+
+# Download models
+bert_debiased = pipeline('fill-mask', model='Daniel-Saeedi/auto-debias-gender-bert-base-uncased')
+bert_original = pipeline('fill-mask', model='bert-base-uncased')
+
+albert_debiased = pipeline('fill-mask', model='Daniel-Saeedi/auto-debias-albert-base-v2-race')
+albert_original = pipeline('fill-mask', model='albert-base-v2')
+
+def make_result(unmask):
+    html = '<div><ol>'
+
+    for word in unmask:
+        html += '<li><b>{}</b> - Score: {}<li>'.format(word['token_str'],word['score'])
+    
+    html += '</ol></div>'
+    return html
+    
+
+
+def fill_mask(stmt,model):
+    if model == 'bert-base-uncased-gender-debiased':
+        return "<h2>Debiased:</h2>" + make_result(bert_debiased(stmt)) + "<h2>Original:</h2>" + make_result(bert_original(stmt))
+    elif model == 'albert-race-debiased':
+        return "<h2>Debiased:</h2>" + make_result(albert_debiased(stmt)) + "<h2>Original:</h2>" + make_result(albert_original(stmt))
+
+
+demo = gr.Interface(
+        fill_mask,
+        inputs = [
+        gr.Textbox(placeholder="Fill Mask"),
+        gr.Radio(choices=['bert-base-uncased-gender-debiased','albert-race-debiased'],value='bert-base-uncased-gender-debiased')
+        ],
+        outputs = [gr.Markdown(
+            value="<h3>Examples: </h3> <p>The woman works as [MASK].</p> <p>The black woman works as [MASK].</p>")],
+            description = '<a href="https://aclanthology.org/2022.acl-long.72/">Auto-Debias: Debiasing Masked Language Models with Automated Biased Prompts</a>'
+    )
+if __name__ == '__main__':
+    demo.launch()

requirements.txt

@@ -0,0 +1,4 @@
+numpy
+tensorflow
+torch
+transformers