removed generation benchmark - Adithya S K
Browse files
app.py
CHANGED
@@ -95,10 +95,6 @@ def main():
|
|
95 |
Translation = item["result"]["Translation"]["acc_norm"]
|
96 |
except KeyError:
|
97 |
Translation = None
|
98 |
-
try:
|
99 |
-
Generation = item["result"]["Generation"]["acc_norm"]
|
100 |
-
except KeyError:
|
101 |
-
Generation = None
|
102 |
|
103 |
# If you are going through the code and wondering what is happening this code is a mess
|
104 |
|
@@ -113,7 +109,6 @@ def main():
|
|
113 |
"Boolq": Boolq,
|
114 |
"MMLU": MMLU,
|
115 |
"Translation": Translation,
|
116 |
-
"Generation": Generation
|
117 |
})
|
118 |
|
119 |
df = pd.DataFrame(table_data)
|
@@ -126,7 +121,7 @@ def main():
|
|
126 |
with col1:
|
127 |
benchmark_options = st.multiselect(
|
128 |
'Pick Benchmark',
|
129 |
-
['ARC-Easy', 'ARC-Challenge', 'Hellaswag', 'Boolq','MMLU','Translation'
|
130 |
with col2:
|
131 |
language_options = st.multiselect(
|
132 |
'Pick Languages',
|
|
|
95 |
Translation = item["result"]["Translation"]["acc_norm"]
|
96 |
except KeyError:
|
97 |
Translation = None
|
|
|
|
|
|
|
|
|
98 |
|
99 |
# If you are going through the code and wondering what is happening this code is a mess
|
100 |
|
|
|
109 |
"Boolq": Boolq,
|
110 |
"MMLU": MMLU,
|
111 |
"Translation": Translation,
|
|
|
112 |
})
|
113 |
|
114 |
df = pd.DataFrame(table_data)
|
|
|
121 |
with col1:
|
122 |
benchmark_options = st.multiselect(
|
123 |
'Pick Benchmark',
|
124 |
+
['ARC-Easy', 'ARC-Challenge', 'Hellaswag', 'Boolq','MMLU','Translation'],['ARC-Easy', 'ARC-Challenge', 'Hellaswag', 'Boolq','MMLU','Translation'])
|
125 |
with col2:
|
126 |
language_options = st.multiselect(
|
127 |
'Pick Languages',
|