Spaces:
Running
Running
benediktstroebl
commited on
Commit
·
bf0e375
1
Parent(s):
387c612
fixed sorting. Modified axis labels
Browse files
app.py
CHANGED
@@ -23,7 +23,7 @@ with gr.Blocks() as demo:
|
|
23 |
with gr.Tab("SWE-Bench"):
|
24 |
with gr.Row():
|
25 |
with gr.Column(scale=1):
|
26 |
-
scatter_plot = gr.Plot(create_scatter_plot(parse_json_files(os.path.join(abs_path, "evals"), 'swebench_lite'), "results_total_cost", "results_accuracy", "Cost", "Accuracy", ["agent_name"]))
|
27 |
with gr.Column(scale=1):
|
28 |
Leaderboard(
|
29 |
value=parse_json_files(os.path.join(abs_path, "evals"), 'swebench_lite'),
|
|
|
23 |
with gr.Tab("SWE-Bench"):
|
24 |
with gr.Row():
|
25 |
with gr.Column(scale=1):
|
26 |
+
scatter_plot = gr.Plot(create_scatter_plot(parse_json_files(os.path.join(abs_path, "evals"), 'swebench_lite'), "results_total_cost", "results_accuracy", "Cost (in USD)", "Accuracy", ["agent_name"]))
|
27 |
with gr.Column(scale=1):
|
28 |
Leaderboard(
|
29 |
value=parse_json_files(os.path.join(abs_path, "evals"), 'swebench_lite'),
|
utils.py
CHANGED
@@ -41,8 +41,13 @@ def parse_json_files(folder_path, benchmark_name):
|
|
41 |
|
42 |
# Create DataFrame from the list of dictionaries
|
43 |
df = pd.DataFrame(data_list)
|
|
|
44 |
|
45 |
-
|
|
|
|
|
|
|
|
|
46 |
|
47 |
|
48 |
def create_scatter_plot(df, x: str, y: str, x_label: str = None, y_label: str = None, hover_data: list = None):
|
|
|
41 |
|
42 |
# Create DataFrame from the list of dictionaries
|
43 |
df = pd.DataFrame(data_list)
|
44 |
+
df = df[df['benchmark_name'] == benchmark_name]
|
45 |
|
46 |
+
# sort df by descending accuracy
|
47 |
+
df = df.sort_values(by='results_accuracy', ascending=False)
|
48 |
+
|
49 |
+
|
50 |
+
return df
|
51 |
|
52 |
|
53 |
def create_scatter_plot(df, x: str, y: str, x_label: str = None, y_label: str = None, hover_data: list = None):
|