Clean up description. Higher rez image

app.py

@@ -210,9 +210,14 @@ demo = gr.Blocks()
 with demo:
     input_texts = gr.Variable([])
     gr.Markdown("**Detect Task Specification at Inference-time.**")
-    gr.Markdown("""Well-specified tasks should have a lower specification metric value. 
-                For example, with a close read, you can see that only Winogender schema sentence numbers (3) and (4) are well-specified:
-                the masked pronoun is coreferent with the `man` or `woman`, for the gendered pronoun resolution task, but the remainder are unspecfied.
+    gr.Markdown("""This method exploits the specification-induced spurious correlations demonstrated in this 
+                [Spurious Correlations Hugging Face Space](https://huggingface.co/spaces/anonymousauthorsanonymous/spurious) to detect task specification at inference-time. 
+                For this method, well-specified tasks should have a lower specification metric value, and unspecified tasks should have a higher specification metric value.
+                """)
+
+    gr.Markdown("""As an example, see the figure below with test sentences from the [Winogender schema](https://aclanthology.org/N18-2002/) for the occupation of `Doctor`. 
+                With a close read, you can see that only sentence numbers (3) and (4) are well-specified for the gendered pronoun resolution task:
+                the masked pronoun is coreferent with the `man` or `woman`; the remainder are unspecfied: the masked pronoun is coreferent with a gender-unspecified person.
                              
                 In this example we have 100\% accurate detection with the specification metric near zero for only sentence (3) and (4).
                 <p align="center">
@@ -221,14 +226,14 @@ with demo:
                 """)
     
 
-    gr.Markdown("**Follow the numbered steps below to test one of the pre-loaded options.** Once you get the hang of it, you can load a new model and/or provide your own input texts.")
+    gr.Markdown("**To test this for yourself, follow the numbered steps below to test one of the pre-loaded options.** Once you get the hang of it, you can load a new model and/or provide your own input texts.")
     gr.Markdown(f"""1) Pick a preloaded BERT-like model.  
         *Note: RoBERTa-large performance is best.*
     2) Pick an Occupation type from the Winogender Schemas evaluation set.   
         *Or select '{PICK_YOUR_OWN_LABEL}' (it need not be about an occupation).*
-    3) Click button to load input texts.   
+    3) Click the first button to load input texts.   
         *Read the sentences to determine which two are well-specified for gendered pronoun coreference resolution. The rest are gender-unspecified.*
-    4) Click button to get Task Specification Metric results!   
+    4) Click the second button to get Task Specification Metric results.
     """)
     
 
@@ -272,7 +277,7 @@ with demo:
     with gr.Row():
         uncertain_btn = gr.Button("4) Click to get Task Specification Metric results!")
     gr.Markdown(
-        """We expect a lower specification metric for well-specified tasks.
+        """We expect a lower specification metric value for well-specified tasks.
         
         Note: If there is an * by a sentence number, then at least one top prediction for that sentence was non-gendered.""")