cutoff 6.04

arena_elo/elo_rating/elo_analysis.py

@@ -171,20 +171,20 @@ def visualize_pairwise_win_fraction(battles, model_order):
         xaxis_side="top",
         title_y=0.07,
         title_x=0.5,
-        # xaxis=dict(
-        #     tickfont=dict(size=16),
-        #     title=dict(font=dict(size=16)),
-        # ),
-        # yaxis=dict(
-        #     tickfont=dict(size=16),
-        #     title=dict(font=dict(size=16)),
-        # ),
+        xaxis=dict(
+            tickfont=dict(size=16),
+            title=dict(font=dict(size=16)),
+        ),
+        yaxis=dict(
+            tickfont=dict(size=16),
+            title=dict(font=dict(size=16)),
+        ),
     )
     fig.update_traces(
-        # textfont=dict(size=16),
-        # colorbar=dict(
-        #     title=dict(font=dict(size=16))
-        # ),
+        textfont=dict(size=16),
+        colorbar=dict(
+            title=dict(font=dict(size=16))
+        ),
         hovertemplate="Model A: %{y}<br>Model B: %{x}<br>Fraction of A Wins: %{z}<extra></extra>"
     )
 
@@ -379,11 +379,11 @@ if __name__ == "__main__":
     
     
     # # save heatmap results in the same directory of the cleaned battle file
-    # win_fraction_heatmap_file = args.clean_battle_file.replace(".json", "_win_fraction_heatmap.jpg")
+    win_fraction_heatmap_file = args.clean_battle_file.replace(".json", "_win_fraction_heatmap.jpg")
     # battle_count_heatmap_file = args.clean_battle_file.replace(".json", "_battle_count_heatmap.jpg")
     # average_win_rate_bar_file = args.clean_battle_file.replace(".json", "_average_win_rate_bar.jpg")
     # bootstrap_elo_rating_file = args.clean_battle_file.replace(".json", "_bootstrap_elo_rating.jpg")
-    # anony_results["win_fraction_heatmap"].write_image(win_fraction_heatmap_file)
+    anony_results["win_fraction_heatmap"].write_image(win_fraction_heatmap_file)
     # anony_results["battle_count_heatmap"].write_image(battle_count_heatmap_file)
     # anony_results["average_win_rate_bar"].write_image(average_win_rate_bar_file)
     # anony_results["bootstrap_elo_rating"].write_image(bootstrap_elo_rating_file)

arena_elo/results/20240602/elo_results_image_editing.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f987c276a9f91055a204fee98a63f32895c94264e3a4fcc41b84776592763e6f
-size 62362
+oid sha256:51cebad92fed9a2c15871c3dae4ef1ffbbaea02b5906f3100e5e9f1dc25e97da
+size 62732

arena_elo/results/20240604/elo_results_t2i_generation.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:62c6b2be4057410cea279902889932cfab9457d9b9e8da3c9cb3f9a37971f9df
-size 68024
+oid sha256:787326f9d085b668bdf7d2f24439f7080a402c7333423f4856c86ef1c11288b8
+size 68373

arena_elo/results/20240604/elo_results_video_generation.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0abd4fd48f01aca7976d4517a31bf755c204dede15a3fab3bbf7b20c064b9181
-size 59813
+oid sha256:e67b05d84d6bccdaefaba734220e3a87d1ef897a4969f415a8a94831a2a9d9a1
+size 60162

arena_elo/results/latest/elo_results_image_editing.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f987c276a9f91055a204fee98a63f32895c94264e3a4fcc41b84776592763e6f
-size 62362
+oid sha256:51cebad92fed9a2c15871c3dae4ef1ffbbaea02b5906f3100e5e9f1dc25e97da
+size 62732

arena_elo/results/latest/elo_results_t2i_generation.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:62c6b2be4057410cea279902889932cfab9457d9b9e8da3c9cb3f9a37971f9df
-size 68024
+oid sha256:787326f9d085b668bdf7d2f24439f7080a402c7333423f4856c86ef1c11288b8
+size 68373

arena_elo/results/latest/elo_results_video_generation.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0abd4fd48f01aca7976d4517a31bf755c204dede15a3fab3bbf7b20c064b9181
-size 59813
+oid sha256:e67b05d84d6bccdaefaba734220e3a87d1ef897a4969f415a8a94831a2a9d9a1
+size 60162