Spaces:
Running
Running
Update src/leaderboard/read_evals.py
Browse files
src/leaderboard/read_evals.py
CHANGED
@@ -111,7 +111,8 @@ class EvalResult:
|
|
111 |
def to_dict(self):
|
112 |
"""Converts the Eval Result to a dict compatible with our dataframe display"""
|
113 |
# weighted average calculation
|
114 |
-
task_weights = [0.25, 0.25, 0.25, 0.09, 0.09, 0.07]
|
|
|
115 |
average = sum(np.array([v for v in self.results.values() if v is not None]) * np.array(task_weights)) / sum(task_weights)
|
116 |
data_dict = {
|
117 |
"eval_name": self.eval_name, # not a column, just a save name,
|
|
|
111 |
def to_dict(self):
|
112 |
"""Converts the Eval Result to a dict compatible with our dataframe display"""
|
113 |
# weighted average calculation
|
114 |
+
#task_weights = [0.25, 0.25, 0.25, 0.09, 0.09, 0.07]
|
115 |
+
task_weights = [0.34, 0.33, 0.33]
|
116 |
average = sum(np.array([v for v in self.results.values() if v is not None]) * np.array(task_weights)) / sum(task_weights)
|
117 |
data_dict = {
|
118 |
"eval_name": self.eval_name, # not a column, just a save name,
|