separate txt and vision avgs

src/leaderboard/read_evals.py

@@ -8,7 +8,7 @@ import dateutil
 import numpy as np
 
 from src.display.formatting import make_clickable_model
-from src.display.utils import AutoEvalColumn, Tasks, TasksMultimodal
+from src.display.utils import AutoEvalColumn, AutoEvalColumnMultimodal, Tasks, TasksMultimodal
 from src.submission.check_validity import is_model_on_hub
 
 
@@ -100,20 +100,21 @@ class EvalResult:
 
     def to_dict(self):
         """Converts the Eval Result to a dict compatible with our dataframe display"""
+        eval_column = AutoEvalColumnMultimodal if self.track.lower() == "multimodal" else AutoEvalColumn
         vision_tasks = ("VQA", "Winoground", "DevBench")
         text_average = sum([v for k, v in self.results.items() if v is not None and k not in vision_tasks]) / len(Tasks)
         data_dict = {
             "eval_name": self.eval_name,  # not a column, just a save name,
-            AutoEvalColumn.model.name: make_clickable_model(self.full_model),
-            AutoEvalColumn.revision.name: self.revision,
-            AutoEvalColumn.text_average.name: text_average,
-            AutoEvalColumn.still_on_hub.name: self.still_on_hub,
+            eval_column.model.name: make_clickable_model(self.full_model),
+            eval_column.revision.name: self.revision,
+            eval_column.text_average.name: text_average,
+            eval_column.still_on_hub.name: self.still_on_hub,
         }
 
         if self.track.lower() == "multimodal":
             taskset = TasksMultimodal
             vision_average = sum([v for k, v in self.results.items() if v is not None and k in vision_tasks]) / len(Tasks)
-            data_dict[AutoEvalColumn.vision_average.name] = vision_average
+            data_dict[eval_column.vision_average.name] = vision_average
         else:
             taskset = Tasks
         for task in taskset:

src/populate.py

@@ -4,7 +4,7 @@ import os
 import pandas as pd
 
 from src.display.formatting import has_no_nan_values, make_clickable_model
-from src.display.utils import AutoEvalColumn, EvalQueueColumn
+from src.display.utils import AutoEvalColumn, AutoEvalColumnMultimodal, EvalQueueColumn
 from src.leaderboard.read_evals import get_raw_eval_results
 
 
@@ -23,7 +23,7 @@ def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchm
 
     df = pd.DataFrame.from_records(all_data_json)
     print(df)
-    df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
+    df = df.sort_values(by=[AutoEvalColumn.text_average.name], ascending=False)
     df = df[cols].round(decimals=1)
 
     # filter out if any of the benchmarks have not been produced