Lisa Dunlap
commited on
Commit
β’
0ba05dc
1
Parent(s):
f61ae52
added arrow colors
Browse files
app.py
CHANGED
@@ -426,6 +426,8 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
|
|
426 |
elo_subset_results = category_elo_results[category]
|
427 |
arena_df = arena_dfs["Total"]
|
428 |
arena_values = get_arena_table(arena_df, model_table_df, arena_subset_df)
|
|
|
|
|
429 |
p1 = elo_subset_results["win_fraction_heatmap"]
|
430 |
p2 = elo_subset_results["battle_count_heatmap"]
|
431 |
p3 = elo_subset_results["bootstrap_elo_rating"]
|
@@ -434,6 +436,18 @@ def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=Fa
|
|
434 |
"""
|
435 |
leaderboard_md = make_category_arena_leaderboard_md(arena_df, arena_subset_df, name=category)
|
436 |
return arena_values, p1, p2, p3, p4, more_stats_md, leaderboard_md
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
437 |
|
438 |
category_dropdown.change(update_leaderboard_and_plots, inputs=[category_dropdown], outputs=[elo_display_df, plot_1, plot_2, plot_3, plot_4, more_stats_md, category_deets])
|
439 |
|
|
|
426 |
elo_subset_results = category_elo_results[category]
|
427 |
arena_df = arena_dfs["Total"]
|
428 |
arena_values = get_arena_table(arena_df, model_table_df, arena_subset_df)
|
429 |
+
if category != "Total":
|
430 |
+
arena_values = update_leaderboard_df(arena_values)
|
431 |
p1 = elo_subset_results["win_fraction_heatmap"]
|
432 |
p2 = elo_subset_results["battle_count_heatmap"]
|
433 |
p3 = elo_subset_results["bootstrap_elo_rating"]
|
|
|
436 |
"""
|
437 |
leaderboard_md = make_category_arena_leaderboard_md(arena_df, arena_subset_df, name=category)
|
438 |
return arena_values, p1, p2, p3, p4, more_stats_md, leaderboard_md
|
439 |
+
|
440 |
+
def update_leaderboard_df(arena_table_vals):
|
441 |
+
elo_datarame = pd.DataFrame(arena_table_vals, columns=["Rank", "Model", "Arena Elo", "95% CI", "Votes", "Organization", "License", "Knowledge Cutoff"])
|
442 |
+
|
443 |
+
# goal: color the rows based on the rank with styler
|
444 |
+
def highlight_max(s):
|
445 |
+
# all items in S which contain up arrow should be green, down arrow should be red, otherwise black
|
446 |
+
return ["color: green" if "\u2191" in v else "color: red" if "\u2193" in v else "" for v in s]
|
447 |
+
|
448 |
+
styled_df = elo_datarame.style.apply(highlight_max, subset=["Rank"])
|
449 |
+
|
450 |
+
return styled_df
|
451 |
|
452 |
category_dropdown.change(update_leaderboard_and_plots, inputs=[category_dropdown], outputs=[elo_display_df, plot_1, plot_2, plot_3, plot_4, more_stats_md, category_deets])
|
453 |
|