Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -33,7 +33,15 @@ def prediction_analyse(prediction_content):
|
|
| 33 |
|
| 34 |
# 遍历 predictions,计算每个 question_type_id 的正确预测数和总预测数
|
| 35 |
for prediction in predictions:
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
question_id = prediction["question_id"]
|
| 38 |
gt_item = ground_truth[question_id]
|
| 39 |
question_type_id = gt_item["question_type_id"]
|
|
@@ -77,10 +85,16 @@ def add_new_eval(
|
|
| 77 |
LLM_name = LLM_name_textbox
|
| 78 |
else:
|
| 79 |
LLM_name = LLM_type
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 80 |
# add new data
|
| 81 |
new_data = [
|
| 82 |
model_type,
|
| 83 |
-
|
| 84 |
LLM_name,
|
| 85 |
each_task_accuracy[1],
|
| 86 |
each_task_accuracy[2],
|
|
@@ -148,7 +162,6 @@ with block:
|
|
| 148 |
)
|
| 149 |
|
| 150 |
def on_checkbox_group_change(selected_columns):
|
| 151 |
-
print("TEst")
|
| 152 |
# pdb.set_trace()
|
| 153 |
selected_columns = [item for item in TASK_INFO if item in selected_columns]
|
| 154 |
present_columns = MODEL_INFO + selected_columns
|
|
@@ -156,8 +169,6 @@ with block:
|
|
| 156 |
updated_headers = present_columns
|
| 157 |
update_datatype = [DATA_TITILE_TYPE[COLUMN_NAMES.index(x)] for x in updated_headers]
|
| 158 |
|
| 159 |
-
# data_component.update(value=updated_data, headers=updated_headers, datatype=update_datatype)
|
| 160 |
-
|
| 161 |
filter_component = gr.components.Dataframe(
|
| 162 |
value=updated_data,
|
| 163 |
headers=updated_headers,
|
|
@@ -202,7 +213,7 @@ with block:
|
|
| 202 |
],
|
| 203 |
label="Model type",
|
| 204 |
multiselect=False,
|
| 205 |
-
value="
|
| 206 |
interactive=True,
|
| 207 |
)
|
| 208 |
model_link = gr.Textbox(
|
|
|
|
| 33 |
|
| 34 |
# 遍历 predictions,计算每个 question_type_id 的正确预测数和总预测数
|
| 35 |
for prediction in predictions:
|
| 36 |
+
# pdb.set_trace()
|
| 37 |
+
prediction = prediction.strip()
|
| 38 |
+
if not prediction:
|
| 39 |
+
continue
|
| 40 |
+
try:
|
| 41 |
+
prediction = json.loads(prediction)
|
| 42 |
+
except json.JSONDecodeError:
|
| 43 |
+
print(f"Warning: Skipping invalid JSON data in line: {prediction}")
|
| 44 |
+
continue
|
| 45 |
question_id = prediction["question_id"]
|
| 46 |
gt_item = ground_truth[question_id]
|
| 47 |
question_type_id = gt_item["question_type_id"]
|
|
|
|
| 85 |
LLM_name = LLM_name_textbox
|
| 86 |
else:
|
| 87 |
LLM_name = LLM_type
|
| 88 |
+
|
| 89 |
+
pdb.set_trace()
|
| 90 |
+
if model_link == '':
|
| 91 |
+
model_name = model_name_textbox # no url
|
| 92 |
+
else:
|
| 93 |
+
model_name = '[' + model_name_textbox + '](' + model_link + ')'
|
| 94 |
# add new data
|
| 95 |
new_data = [
|
| 96 |
model_type,
|
| 97 |
+
model_name,
|
| 98 |
LLM_name,
|
| 99 |
each_task_accuracy[1],
|
| 100 |
each_task_accuracy[2],
|
|
|
|
| 162 |
)
|
| 163 |
|
| 164 |
def on_checkbox_group_change(selected_columns):
|
|
|
|
| 165 |
# pdb.set_trace()
|
| 166 |
selected_columns = [item for item in TASK_INFO if item in selected_columns]
|
| 167 |
present_columns = MODEL_INFO + selected_columns
|
|
|
|
| 169 |
updated_headers = present_columns
|
| 170 |
update_datatype = [DATA_TITILE_TYPE[COLUMN_NAMES.index(x)] for x in updated_headers]
|
| 171 |
|
|
|
|
|
|
|
| 172 |
filter_component = gr.components.Dataframe(
|
| 173 |
value=updated_data,
|
| 174 |
headers=updated_headers,
|
|
|
|
| 213 |
],
|
| 214 |
label="Model type",
|
| 215 |
multiselect=False,
|
| 216 |
+
value=ModelType.FT.to_str(" : "),
|
| 217 |
interactive=True,
|
| 218 |
)
|
| 219 |
model_link = gr.Textbox(
|