Changed background in Accordions and per sentence progress in get sentences

app.py

@@ -275,8 +275,6 @@ def _genSentenceCoverMsg(test_sentences, total_att_terms, bias_spec, isGen=False
     # missing pairs spec
     bt_mgr.genMissingPairsSpec(bias_spec, test_sentences_df)
 
-
-
     att1_missing_num = sum([v for k, v in att1_missing.items()])
     att2_missing_num = sum([v for k, v in att2_missing.items()])
     total_missing = att1_missing_num + att2_missing_num
@@ -287,9 +285,9 @@ def _genSentenceCoverMsg(test_sentences, total_att_terms, bias_spec, isGen=False
     source_msg = "Found" if isGen==False else "Generated"
     if num_covered_atts >= total_att_terms:
         if total_missing > 0:
-            info_msg = f"**{source_msg} {len(test_sentences)} sentences covering all bias specification attributes, but some attributes are underepresented. Generating additional {total_missing} sentences is suggested.**"
+            info_msg = f"**{source_msg} {len(test_sentences)} sentences covering all bias specification attributes, but some attributes are underepresented (see at the bottom). Generating additional {total_missing} sentences is suggested.**"
         else:
-            info_msg = f"**{source_msg} {len(test_sentences)} sentences covering all bias specification attributes. Please select model to test.**"
+            info_msg = f"**{source_msg} {len(test_sentences)} sentences covering all bias specification attributes (see at the bottom). Please select model to test.**"
     else:
         info_msg = f"**{source_msg} {len(test_sentences)} sentences covering {num_covered_atts} of {total_att_terms} attributes. Please select model to test.**"
 
@@ -707,8 +705,7 @@ def useOnlineGen(value):
 
 def changeTerm(evt: gr.EventData):
     global G_CORE_BIAS_NAME
-
-    print("Bias is custom now...")
+    #print("Bias is custom now...")
 
     G_CORE_BIAS_NAME = None
 
@@ -805,7 +802,10 @@ css_adds = "#group_row {background: white; border-color: white;} \
                #filled:hover .tooltiptext_left {visibility: visible;} \
                #empty:hover .tooltiptext_left {visibility: visible;} \
                #filled:hover .tooltiptext_right {visibility: visible;} \
-               #empty:hover .tooltiptext_right {visibility: visible;}"
+               #empty:hover .tooltiptext_right {visibility: visible;} \
+               #add_interpret {border: 1px solid #f1f5f9; background: #F8FAFC;} \
+               #per_sent_bias_accordion {border: 1px solid #f1f5f9; background: #F8FAFC;} \
+               #test_sentences_accordion {border: 1px solid #f1f5f9; background: #F8FAFC;}"
 
 #'bethecloud/storj_theme'
 with gr.Blocks(theme=soft, title="Social Bias Testing in Language Models",
@@ -917,7 +917,7 @@ with gr.Blocks(theme=soft, title="Social Bias Testing in Language Models",
             gr.Markdown("    ")
         
         with gr.Row(visible=False) as row_sentences:
-            with gr.Accordion(label="Test Sentences", open=False, visible=False) as acc_test_sentences:
+            with gr.Accordion(label="Test Sentences", open=False, visible=False, elem_id="test_sentences_accordion") as acc_test_sentences:
                 test_sentences = gr.DataFrame(
                             headers=["Sentence", "Alternative Sentence", "Group term 1", "Group term 2", "Attribute term"],
                             datatype=["str", "str", "str", "str", "str"],
@@ -950,8 +950,8 @@ with gr.Blocks(theme=soft, title="Social Bias Testing in Language Models",
                 model_bias_label = gr.Label(num_top_classes=1, label="% stereotyped choices (↑ more bias)",
                                             elem_id="res_label",
                                             show_label=False)
-                with gr.Accordion("Additional Interpretation", open=False, visible=True):
-                    interpretation_msg = gr.HTML(value="Interpretation: Stereotype Score metric details in <a href='https://arxiv.org/abs/2004.09456'>Nadeem'20<a>", visible=False)
+                with gr.Accordion("Additional Interpretation", open=False, visible=True, elem_id="add_interpret") as interpret_accordion:
+                    interpretation_msg = gr.HTML(value="Interpretation: Stereotype Score metric details in <a href='https://arxiv.org/abs/2004.09456'>Nadeem et al. 2020<a>", visible=False)
 
                 lbl_attrib_bias = gr.Markdown("**Bias in the Context of Attributes** - % stereotyped choices (↑ more bias)")
                 #gr.Markdown("**Legend**")
@@ -971,13 +971,13 @@ with gr.Blocks(theme=soft, title="Social Bias Testing in Language Models",
                         gr.Markdown("#### Attribute Group 2")
                         attribute_bias_html_antistereo = gr.HTML()
             
-                gr.HTML(value="Visualization inspired by <a href='https://www.bloomberg.com/graphics/2023-generative-ai-bias/' target='_blank'>Bloomberg article on bias in text-to-image models</a>.")
+                gr.HTML(value="Visualization inspired by <a href='https://www.bloomberg.com/graphics/2023-generative-ai-bias/' target='_blank'>Bloomberg article on bias in text-to-image models</a>.<br />While we perform several checks, the tool still relies on ChatGPT generations, please examine the sentences for quality.")
                 save_msg = gr.HTML(value="<span style=\"color:black\">Bias test result saved! </span>", 
                                 visible=False)
                 
         with gr.Row():
             with gr.Column(scale=2):
-                with gr.Accordion("Per Sentence Bias Results", open=False, visible=True):
+                with gr.Accordion("Per Sentence Bias Results", open=False, visible=True, elem_id="per_sent_bias_accordion"):
                     test_pairs = gr.DataFrame(
                             headers=["group_term", "template", "att_term_1", "att_term_2","label_1","label_2"],
                             datatype=["str", "str", "str", "str", "str", "str"],
@@ -1071,7 +1071,7 @@ with gr.Blocks(theme=soft, title="Social Bias Testing in Language Models",
     exp_button.click(export_csv, 
                      inputs=[test_pairs, group1, group2, att1, att2], 
                      outputs=[csv])
-
+    
     # Changing any of the bias specification terms
     group1.change(fn=changeTerm, inputs=[], outputs=[csv])
     group2.change(fn=changeTerm, inputs=[], outputs=[csv])

mgr_bias_scoring.py

@@ -905,7 +905,7 @@ def _constructInterpretationMsg(bias_spec, num_sentences, model_name, bias_stats
   interpret_msg += "<div style=\"margin-top: 3px; margin-left: 3px\"><b>◼ </b>" + att1_msg + "<br /></div>"
   interpret_msg += "<div style=\"margin-top: 3px; margin-left: 3px; margin-bottom: 3px\"><b>◼ </b>" + att2_msg + "<br /></div>"
   interpret_msg += "Please examine the exact test sentences used below."
-  interpret_msg += "<br />More details about Stereotype Score metric: <a href='https://arxiv.org/abs/2004.09456' target='_blank'>Nadeem'20<a>"
+  interpret_msg += "<br />More details about Stereotype Score metric: <a href='https://arxiv.org/abs/2004.09456' target='_blank'>Nadeem et al. 2020<a>"
 
   return interpret_msg
   

mgr_requests.py

@@ -171,9 +171,14 @@ def _getSavedSentences(bias_spec, progress, use_paper_sentences):
         att_list.extend(att_list_nospace)
         att_list = list(set(att_list))
 
-        progress(gi/len(g1+g2), desc=f"{g_term}")
+        #progress(gi/len(g1+g2), desc=f"{g_term}")
 
-        _, sentence_df, _ = smgr.getSavedSentences(g_term)
+        _, sentence_df, _ = smgr.getSavedSentences(g_term)#, gi, len(g1+g2), progress)
+        if sentence_df.shape[0] > 0:
+            progress(gi/len(g1+g2), desc=f"{sentence_df['sentence'].tolist()[0]}")
+        else:
+            progress(gi/len(g1+g2), desc=f"{g_term}")
+        
         # only take from paper & gpt3.5
         flt_gen_models = ["gpt-3.5","gpt-3.5-turbo","gpt-4"]
         print(f"Before filter: {sentence_df.shape[0]}")

mgr_sentences.py

@@ -90,13 +90,14 @@ def get_sentence_csv(file_path: str):
   
   return df
 
-def getSavedSentences(grp):
+def getSavedSentences(grp): #, gi, total_grp_len, progress):
     filename = f"{grp.replace(' ','-')}.csv"
     sentence_df = pd.DataFrame()
 
     try:
         text = f"Loading sentences: {filename}\n"
         sentence_df = get_sentence_csv(filename)
+        #progress(gi/total_grp_len, desc=f"{sentence_df[0]}")
 
     except Exception as e:
         text = f"Error, no saved generations for {filename}"