Update text content

app.py

@@ -80,7 +80,7 @@ def create_species_guide_content():
         gr.Image("species_id/jaguar.jpg", label="Jaguar example image", show_label=False)
 
         gr.Markdown("""
-        The largest cat in the Americas, with a stocky, muscular build and a broad head; its golden coat is patterned with rosettes that often have central spots inside.
+        #### The largest cat in the Americas, with a stocky, muscular build and a broad head. Coat is patterned with rosettes that often have central spots inside.
 
         ----
 
@@ -91,7 +91,7 @@ def create_species_guide_content():
         gr.Image("species_id/ocelot.jpg", label="Ocelot example image", show_label=False)
 
         gr.Markdown("""
-        A medium-sized spotted cat about twice the size of a domestic cat, with a slender body, large eyes, and striking chain-link or stripe-like rosettes. It differs from jaguars by its smaller size, leaner build, and more elongated markings.
+        #### Smaller and leaner than a jaguar, with more elongated markings and rounder ears.
 
         ----
 
@@ -101,7 +101,7 @@ def create_species_guide_content():
         gr.Image("species_id/mountainlion.jpg", label="Mountain lion example image", show_label=False)
 
         gr.Markdown("""
-        Also called cougar or puma, this cat has a plain tawny or grayish coat without spots or rosettes. Its long tail and uniformly colored fur distinguish it from jaguars and ocelots.
+        #### Also called cougar or puma, this cat has a plain tawny or grayish coat without spots. Its long tail and uniformly colored fur distinguish it from jaguars and ocelots.
 
         ----
 
@@ -112,7 +112,7 @@ def create_species_guide_content():
         gr.Image("species_id/commoneland.jpg", label="Eland example image", show_label=False)
 
         gr.Markdown("""
-        The largest antelope species, with a robust body, spiraled horns on both sexes, and a characteristic dewlap hanging from the throat. It differs from waterbuck by its lighter tan coat, faint body stripes, and massive size.
+        ### The largest antelope species. Identifiable by its spiraled horns on both sexes. Lighter tan coat than a waterbuck.
 
         ----
 
@@ -122,7 +122,7 @@ def create_species_guide_content():
         gr.Image("species_id/waterbuck.jpg", label="Waterbuck example image", show_label=False)
 
         gr.Markdown("""
-        A shaggy, dark brown antelope recognized by its white rump ring and backward-curving horns in males. Smaller and darker than the common eland, waterbuck prefer wet habitats and lack the eland's throat dewlap.
+        #### A shaggy, dark brown antelope. Identifiable by backward-curving horns in males, no horns on females. Larger, rounder ears and darker coat than the common eland.
 
         ----
 
@@ -837,20 +837,32 @@ with gr.Blocks(title="CODA: Wildlife Photo Classification Challenge",
             gr.Markdown("""
             ## CODA Model Selection Probabilities
 
-            This chart shows CODA's current confidence in each candidate model being the best performer.
+            This chart shows CODA's current confidence in each candidate classifier being the best performer.
 
             **How to read this chart:**
-            - Each bar represents one of the candidate machine learning models
+            - Each bar represents one of the candidate machine learning classifiers
             - The height of each bar shows the probability (0-100%) that this model is the best, according to CODA
             - The orange bar indicates CODA's current best guess
             - As you provide more labels, CODA updates these probabilities
 
             **What you'll see:**
-            - CODA initializes these probabilities based on each model's agreement with the consensus, providing informative priors
+            - CODA initializes these probabilities based on each classifier's agreement with the consensus votes of *all* classifiers, 
+                        providing informative priors
             - As you label images, some models will gain confidence while others lose it
             - The goal is for one model to clearly emerge as the winner
                         
+            More details can be found in [the paper](https://www.arxiv.org/abs/2507.23771)!
+            
+            **What models are these?**
+                        
+            For this demo, we selected 5 zero-shot classifiers that would be reasonable choices for someone who 
+                        wanted to classify wildlife imagery. The models are: facebook/PE-Core-L14-336, 
+                        google/siglip2-so400m-patch16-naflex, openai/clip-vit-large-patch14, imageomics/bioclip-2, and
+                        laion/CLIP-ViT-L-14-laion2B-s32B-b82K. Our goal is not to make any general claims about the performance
+                        of these models but rather to provide a realistic set of candidates for demonstrating CODA.
+                        
             """)
+            gr.HTML("<div style='margin-top: 0.1em;'></div>")
             prob_help_close = gr.Button("Close", variant="secondary")
 
     with gr.Group(visible=False, elem_classes="help-popup-overlay") as acc_help_popup:
@@ -877,10 +889,27 @@ with gr.Blocks(title="CODA: Wildlife Photo Classification Challenge",
     with gr.Group(visible=False, elem_classes="help-popup-overlay") as selection_help_popup:
         with gr.Group(elem_classes="help-popup-content"):
             gr.Markdown("""
-            ## How CODA Selects Images for Labeling
+            ## How CODA selects images for labeling
+
+            CODA selects images that best differentiate top-performing classifiers from each other. It
+                        does this by constructing a probabilistic model of which classifier is best (see
+                        the plot at the bottom-left). Each iteration, CODA selects an image to be labeled
+                        based on how much a label for that image is expected to affect the probabilistic model. 
+
+            Intuitively, CODA will select images where the top classifiers disagree, since knowing the ground truth for these images will provide 
+                        the most information about which classifier is best overall.
+                        
+            More details can be found in [the paper](https://www.arxiv.org/abs/2507.23771)!
+                        
+            **What data is this?**
+                        
+            We selected a subset of 5 species from the iWildcam dataset, and subsampled a dataset of ~500 images for this demo.
+                        Each refresh will generate a slightly different subset, leading to slightly different model selection 
+                        performance.
 
-            [Placeholder]
             """)
+            gr.HTML("<div style='margin-top: 0.1em;'></div>")
+
             selection_help_close = gr.Button("Close", variant="secondary")
 
     # Species guide popup during demo