remove duplicates and change colors

app.py

@@ -17,7 +17,8 @@ def draw_leaderboard():
     st.markdown(PAGE_INFO)
     st.subheader("Evaluation results:")
     st.text('Each tab corresponds to a task, avg - averaged scores over qa1-5 tasks.')
-    st.markdown('Predictions of all evaluated models: [BABILong evals](https://huggingface.co/datasets/RMT-team/babilong_evals)')
+    st.markdown('Predictions of all evaluated models: '
+                '[BABILong evals](https://huggingface.co/datasets/RMT-team/babilong_evals)')
     search_term = st.text_input("Search models:", "")
 
     tabs = st.tabs([str(task) for task in tasks])
@@ -32,10 +33,10 @@ def draw_leaderboard():
             row_height = 35
             height = (len(task_df) + 1) * row_height
 
-            styled = style_dataframe(task_df).format(precision=1)
+            styled_df = style_dataframe(task_df).format(precision=1)
 
             st.dataframe(
-                styled,
+                styled_df,
                 width=1030,
                 height=height,
             )

draw_utils.py

@@ -1,5 +1,6 @@
 import pandas as pd
 import numpy as np
+from matplotlib.colors import LinearSegmentedColormap
 
 PAGE_MARKDOWN = """
  <style>
@@ -26,7 +27,19 @@ def load_results():
     old_results = pd.read_csv(old_results_path)
     new_results = pd.read_csv(new_results_path)
 
-    res = pd.concat([old_results, new_results])
+    def normalize_model_name(name):
+        if '/' in name:
+            name = name.split('/')[-1]
+        return name.lower()
+
+    old_results['normalized_name'] = old_results['model_name'].apply(normalize_model_name)
+    new_results['normalized_name'] = new_results['model_name'].apply(normalize_model_name)
+    # clean duplicate models in v0 results and new results
+    duplicate_models = set(old_results['normalized_name']).intersection(set(new_results['normalized_name']))
+    old_results_filtered = old_results[~old_results['normalized_name'].isin(duplicate_models)]
+    res = pd.concat([old_results_filtered, new_results])
+    res.drop('normalized_name', axis=1, inplace=True)
+
     res.replace(-1, np.nan, inplace=True)
     res['<=32k'] = res[LENGTHS_32k].mean(axis=1)
     res['<=128k'] = res[LENGTHS_128k].mean(axis=1)
@@ -42,6 +55,30 @@ def load_results():
     return res
 
 
+# from pandas/io/formats/style.py
+def relative_luminance(rgba) -> float:
+    """
+    Calculate relative luminance of a color.
+
+    The calculation adheres to the W3C standards
+    (https://www.w3.org/WAI/GL/wiki/Relative_luminance)
+
+    Parameters
+    ----------
+    color : rgb or rgba tuple
+
+    Returns
+    -------
+    float
+        The relative luminance as a value from 0 to 1
+    """
+    r, g, b = (
+        x / 12.92 if x <= 0.04045 else ((x + 0.055) / 1.055) ** 2.4
+        for x in rgba[:3]
+    )
+    return 0.2126 * r + 0.7152 * g + 0.0722 * b
+
+
 def style_dataframe(df):
     """
     Style a pandas DataFrame with a color gradient.
@@ -50,15 +87,17 @@ def style_dataframe(df):
     numeric_columns = styled_df.columns[1:]
 
     def color_scale(val):
+        cmap = LinearSegmentedColormap.from_list('ryg', ["red", "yellow", "green"], N=256)
+
         if pd.isna(val):
-            return 'background-color: white; color: white'
+            return 'background-color: white; color: white;'
+
         min_val = 0
         max_val = 100
-        normalized = (val - min_val) / (max_val - min_val) if max_val > min_val else 0.5
-        r = int(255 * (1 - normalized) + 144 * normalized)
-        g = int(204 * (1 - normalized) + 238 * normalized)
-        b = int(204 * (1 - normalized) + 180 * normalized)
-        return f'background-color: rgb({r}, {g}, {b})'
-
-    styled = styled_df.style.map(color_scale, subset=numeric_columns)
-    return styled
+        normalized = (val - min_val) / (max_val - min_val)
+        rgba = cmap(normalized)
+        text_color = 'white' if relative_luminance(rgba) < 0.408 else 'black'
+        return f'background-color: rgba({rgba[0]*255},{rgba[1]*255},{rgba[2]*255},{rgba[3]}); color: {text_color}'
+
+    styled_df = styled_df.style.map(color_scale, subset=numeric_columns)
+    return styled_df