OpenCompass/opencompass/datasets/calm/evaluation/labeling/CLADDER.py
Peng Bo 07c96ac659
Calm dataset (#1385)
* Add CALM Dataset
2024-08-01 10:03:21 +08:00

59 lines
2.2 KiB
Python

# flake8: noqa: E501
from .common_answers import (common_false_list, common_start_false_dict,
common_start_true_dict, common_true_list)
def get_gt_label(item):
if item['gt_answer'] == 'yes':
gt_label = 1
elif item['gt_answer'] == 'no':
gt_label = 0
return gt_label
def get_pred_label(model_response, item, prompt_style, type):
model_response = model_response.strip().lower()
low_index = len(model_response)
start_str1_dict = common_start_true_dict
start_str2_dict = common_start_false_dict
start_option1_list, start_option2_list = [], []
# some of the model will give response containing the question,
# we usually preprocess the response to remove the question part,
# but sometimes due to the model's response format, some of the
# question part is not removed, so here we are checking the
# response with the question part as well.
for key in start_str1_dict.keys():
for str1 in start_str1_dict[key]:
for i in range(key, len(str1) + 1):
start_option1_list.append(str1[-i:])
for key in start_str2_dict.keys():
for str2 in start_str2_dict[key]:
for i in range(key, len(str2) + 1):
start_option2_list.append(str2[-i:])
inner_option1_list = ['method 1 is more correct', '使用方法1更准确'
] + common_true_list
inner_option2_list = [
'method 2 is more correct', 'method 2 is correct',
'correct to use method 2', '方法2比方法1更准确', '方法2'
] + common_false_list
if model_response.startswith(tuple(start_option1_list)):
label = 1
elif model_response.startswith(tuple(start_option2_list)):
label = 0
elif any(
model_response.find(option) > -1 and
(low_index := min(low_index, model_response.find(option))) > -1
for option in inner_option1_list):
label = 1
if any(option in model_response
and model_response.find(option) < low_index
for option in inner_option2_list):
label = 0
elif any(response in model_response for response in inner_option2_list):
label = 0
else:
return -1
return label