Spaces:
Running
Running
zhimin-z
commited on
Commit
·
bfeb8ad
1
Parent(s):
de0f977
add
Browse files
app.py
CHANGED
|
@@ -406,7 +406,7 @@ def get_leaderboard_data(vote_entry=None, use_cache=True):
|
|
| 406 |
"Elo Score": 2,
|
| 407 |
"Conversation Efficiency Index": 2,
|
| 408 |
"Consistency Score": 2,
|
| 409 |
-
"
|
| 410 |
"Bradley-Terry Coefficient": 2,
|
| 411 |
"Eigenvector Centrality Value": 2,
|
| 412 |
"Newman Modularity Score": 2,
|
|
@@ -444,7 +444,7 @@ def get_leaderboard_data(vote_entry=None, use_cache=True):
|
|
| 444 |
"Elo Score",
|
| 445 |
"Conversation Efficiency Index",
|
| 446 |
"Consistency Score",
|
| 447 |
-
"
|
| 448 |
"Bradley-Terry Coefficient",
|
| 449 |
"Eigenvector Centrality Value",
|
| 450 |
"Newman Modularity Score",
|
|
@@ -586,7 +586,7 @@ def get_leaderboard_data(vote_entry=None, use_cache=True):
|
|
| 586 |
"Elo Score": elo_scores.values,
|
| 587 |
"Conversation Efficiency Index": cei_result.values,
|
| 588 |
"Consistency Score": mcs_result.values,
|
| 589 |
-
"
|
| 590 |
"Bradley-Terry Coefficient": bt_scores.values,
|
| 591 |
"Eigenvector Centrality Value": eigen_scores.values,
|
| 592 |
"Newman Modularity Score": newman_scores.values,
|
|
@@ -598,7 +598,7 @@ def get_leaderboard_data(vote_entry=None, use_cache=True):
|
|
| 598 |
leaderboard_data = leaderboard_data.round(
|
| 599 |
{
|
| 600 |
"Elo Score": 2,
|
| 601 |
-
"
|
| 602 |
"Bradley-Terry Coefficient": 2,
|
| 603 |
"Eigenvector Centrality Value": 2,
|
| 604 |
"Newman Modularity Score": 2,
|
|
@@ -744,12 +744,12 @@ with gr.Blocks(title="SWE-Model-Arena", theme=gr.themes.Soft()) as app:
|
|
| 744 |
label="Consistency Score"
|
| 745 |
),
|
| 746 |
ColumnFilter(
|
| 747 |
-
"
|
| 748 |
min=0.0,
|
| 749 |
max=1.0,
|
| 750 |
default=[0.0, 1.0],
|
| 751 |
type="slider",
|
| 752 |
-
label="
|
| 753 |
),
|
| 754 |
ColumnFilter(
|
| 755 |
"Bradley-Terry Coefficient",
|
|
|
|
| 406 |
"Elo Score": 2,
|
| 407 |
"Conversation Efficiency Index": 2,
|
| 408 |
"Consistency Score": 2,
|
| 409 |
+
"Win Rate": 2,
|
| 410 |
"Bradley-Terry Coefficient": 2,
|
| 411 |
"Eigenvector Centrality Value": 2,
|
| 412 |
"Newman Modularity Score": 2,
|
|
|
|
| 444 |
"Elo Score",
|
| 445 |
"Conversation Efficiency Index",
|
| 446 |
"Consistency Score",
|
| 447 |
+
"Win Rate",
|
| 448 |
"Bradley-Terry Coefficient",
|
| 449 |
"Eigenvector Centrality Value",
|
| 450 |
"Newman Modularity Score",
|
|
|
|
| 586 |
"Elo Score": elo_scores.values,
|
| 587 |
"Conversation Efficiency Index": cei_result.values,
|
| 588 |
"Consistency Score": mcs_result.values,
|
| 589 |
+
"Win Rate": avr_scores.values,
|
| 590 |
"Bradley-Terry Coefficient": bt_scores.values,
|
| 591 |
"Eigenvector Centrality Value": eigen_scores.values,
|
| 592 |
"Newman Modularity Score": newman_scores.values,
|
|
|
|
| 598 |
leaderboard_data = leaderboard_data.round(
|
| 599 |
{
|
| 600 |
"Elo Score": 2,
|
| 601 |
+
"Win Rate": 2,
|
| 602 |
"Bradley-Terry Coefficient": 2,
|
| 603 |
"Eigenvector Centrality Value": 2,
|
| 604 |
"Newman Modularity Score": 2,
|
|
|
|
| 744 |
label="Consistency Score"
|
| 745 |
),
|
| 746 |
ColumnFilter(
|
| 747 |
+
"Win Rate",
|
| 748 |
min=0.0,
|
| 749 |
max=1.0,
|
| 750 |
default=[0.0, 1.0],
|
| 751 |
type="slider",
|
| 752 |
+
label="Win Rate"
|
| 753 |
),
|
| 754 |
ColumnFilter(
|
| 755 |
"Bradley-Terry Coefficient",
|