add the "best per language" option

leaderboard.py

@@ -38,8 +38,10 @@ def leaderboard_tab():
     
     metrics = ['metricx_both', 'xcomet_both', 'CHRFpp', 'glotlid_ref']
     systems = sorted(set(stats["system"]))
+    levels = ["sentence_level", "paragraph_level"]
     ALL = "ALL"
     MEAN = "Average"
+    BEST = "Best"
     XX2EN = "Everything-into-English"
     EN2XX = "English-into-Everything"
 
@@ -59,9 +61,7 @@ def leaderboard_tab():
 
         gr.Markdown("## Systems ranking")
         # Inputs
-        gr_level = gr.Dropdown(
-            ["sentence_level", "paragraph_level"], value="sentence_level", label="Level"
-        )
+        gr_level = gr.Dropdown(levels, value="sentence_level", label="Level")
         gr_src_lang = gr.Dropdown([ALL] + sorted(langs_src), value=ALL, label="Source lang")
         gr_tgt_lang = gr.Dropdown([ALL] + sorted(langs_tgt), value=ALL, label="Target lang")
 
@@ -105,12 +105,13 @@ def leaderboard_tab():
         gr_tgt_lang.input(fn=tgt2src, inputs=[gr_src_lang, gr_tgt_lang], outputs=gr_src_lang)
 
         gr.Markdown("## Languages difficulty")
-        gr_system = gr.Dropdown([MEAN] + systems, value=MEAN, label="Translation system")
+        gr_system = gr.Dropdown([MEAN, BEST] + systems, value=MEAN, label="Translation system")
         gr_direction = gr.Dropdown([XX2EN, EN2XX], value=XX2EN, label="Translation direction")
         gr_metric = gr.Dropdown(metrics, label="Quality metric", value="metricx_both")
-        bar_controls = [gr_system, gr_direction, gr_metric]
+        gr_level2 = gr.Dropdown(levels, value="sentence_level", label="Level")
+        bar_controls = [gr_system, gr_direction, gr_metric, gr_level2]
 
-        def get_hist(system, direction, metric):
+        def get_hist(system, direction, metric, level):
             # decide on the data to process
             if direction == EN2XX:
                 direction_filter = stats['src_lang'].eq('eng_Latn')
@@ -118,14 +119,22 @@ def leaderboard_tab():
             else:
                 direction_filter = stats['tgt_lang'].eq('eng_Latn')
                 lang_col = "src_lang"
-            if system == MEAN:
+            if system in (MEAN, BEST):
                 system_filter = stats["system"].astype(bool)
             else:
                 system_filter = stats['system'].eq(system)
-            subset = stats[system_filter & direction_filter]
+            subset = stats[system_filter & direction_filter & stats["level"].eq(level)]
 
             # Compute the means and update the plot
-            means = subset.groupby(lang_col)[metric].mean().sort_values(
+            grouped = subset.groupby(lang_col)[metric]
+            if system == BEST:
+                if metric == "metricx_both":
+                    means = grouped.min()
+                else:
+                    means = grouped.max()
+            else:
+                means = grouped.mean()
+            means = means.sort_values(
                 ascending=(metric=="metricx_both")
             )
             means = means.to_frame().reset_index()
@@ -135,7 +144,7 @@ def leaderboard_tab():
                 sort="y", 
             )
         
-        default_bar = get_hist(gr_system.value, gr_direction.value, gr_metric.value)
+        default_bar = get_hist(*[x.value for x in bar_controls])
         gr_barplot = gr.BarPlot(**default_bar)
 
         for inp in bar_controls: