-
Notifications
You must be signed in to change notification settings - Fork 2
/
gen_dataset.py
283 lines (243 loc) · 12.3 KB
/
gen_dataset.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
import json
from tqdm import tqdm
import os
import argparse
from demo_hub import demo_hub
from prompt_hub import prompt_hub
from utils.agents import Agent
from config import opt
from test import test
class Hub():
prompt_hub = prompt_hub
demo_hub = demo_hub
hub = Hub()
task_description_dict = {'gsm8k': 'Mathematical Reasoning Task', 'strategyqa': 'Commonsense Reasoning Task', 'clutrr': 'Kinship Reasoning Task', 'boolq': 'Reading Comprehension Task'}
def get_args_parser():
parser = argparse.ArgumentParser('Self-Evolving Benchmark', add_help=False)
parser.add_argument('--dataset', default='', type=str,
help='dataset: gsm8k, clutrr, strategyqa, boolq')
parser.add_argument('--mode', default='', type=str,
help='mode: paraphrase, addnoise, reversepolar, alternative, complex, retrieval, planning, knowledge')
parser.add_argument('--data_num', default=100, type=int,
help='the number of data to generate')
parser.add_argument('--local_url', default=None, type=int,
help='url of the local model')
parser.add_argument('--debug', action='store_true',
help='debug mode')
return parser
def output_result(dataset, mode, instance_list):
instance_list = sorted(instance_list, key=lambda x: x['id'])
result_path = os.path.join('datasets', dataset, mode + '.jsonl')
with open(result_path, 'w', encoding='utf-8') as w_f:
for instance in instance_list:
w_f.write(json.dumps(instance, ensure_ascii=False) + '\n')
def get_test_data(args, dataset, mode, data_num=100):
filtered_data_path = os.path.join('results', dataset, mode, f'results_gpt_4_origin.json')
evolved_dataset_path = os.path.join('datasets', dataset, mode + '.jsonl')
if not os.path.exists(filtered_data_path):
print("Filtered data file not found, run GPT-4 to filter data first...")
success = test(args=args, test_data_path=os.path.join('datasets', dataset, 'test.jsonl'), result_path=filtered_data_path, model='gpt_4', dataset=dataset, mode='origin', data_num=data_num)
if not success:
raise ValueError('Original dataset not found, please first set up the original dataset at the path: datasets/{dataset}/test.jsonl')
with open(filtered_data_path, 'r') as r_f:
all_data = json.load(r_f)
exist_ids = []
exist_instances = []
if os.path.exists(evolved_dataset_path):
with open(evolved_dataset_path, 'r', encoding='utf-8') as r_f:
exist_instances = [json.loads(line) for line in r_f]
exist_ids = [each['id'] for each in exist_instances]
return all_data, exist_instances, exist_ids
def to_string(para):
if type(para) == bool:
if para:
return 'Yes'
else:
return 'No'
else:
return para.strip()
def get_problem(each, dataset, mode):
# gsm8k
if dataset == 'gsm8k':
context = ". ".join(each['question'].split(". ")[:-1]).strip() + "."
question = each['question'].split(". ")[-1].strip()
answer = to_string(each['answer'])
return {'context': context, 'question': question, 'answer': answer}
# hotpotqa
elif dataset == 'hotpotqa' or dataset == 'clutrr':
context = each['context'].replace('\n', ' ').strip()
question = each['question'].strip()
answer = each['answer'].strip()
return {'context': context, 'question': question, 'answer': answer}
elif dataset == 'spartqa':
context = each['context'].replace('\n', ' ').strip()
question = each['question'].replace('\n', ' ').strip()
answer = ', '.join(each['answer'])
return {'context': context, 'question': question, 'answer': answer}
# strategyqa
elif dataset == 'strategyqa':
context = None
question = to_string(each['question'])
answer = to_string(each['answer'])
return {'context': context, 'question': question, 'answer': answer}
elif dataset == 'boolq':
context = each['context']
question = each['question'] if mode != 'complex' else None
answer = to_string(each['answer']) if mode != 'complex' else None
return {'context': context, 'question': question, 'answer': answer}
def construct_new_inst(new_inst, status=None, context=None, question=None, answer=None, option=None, cause=None):
if status:
new_inst['status'] = status
if context:
new_inst['context'] = context
if question:
new_inst['question'] = question
if answer:
new_inst['answer'] = answer
if cause:
new_inst['cause'] = cause
if option:
new_inst['option'] = option
return new_inst
def construct_new_inst_from_response(new_inst, problem, response_list, mode):
if mode == 'paraphrase' or mode == 'addnoise':
assert "context: " in response_list[0].lower()
new_context = response_list[0].lower().split("context: ")[-1].strip()
return construct_new_inst(new_inst, context=new_context, question=problem["question"], answer=problem["answer"])
elif mode == 'reversepolar':
if "context: " not in response_list[0].lower() or "answer: " not in response_list[-1].lower():
return construct_new_inst(new_inst, status='fail', cause='Generator error: no context or answer')
new_context = response_list[0].lower().split("context: ")[-1].strip()
new_answer = response_list[-1].lower().split("answer: ")[-1].strip()
return construct_new_inst(new_inst, context=new_context, question=problem["question"], answer=new_answer)
elif mode == 'alternative' or mode == 'complex' or mode in opt.sub_ability_mode:
assert "question: " in response_list[0].lower() and "answer: " in response_list[-1].lower()
new_question = response_list[0].lower().strip().split("question: ")[1].strip()
new_answer = response_list[-1].lower().strip().split("answer: ")[1].strip()
return construct_new_inst(new_inst, context=problem["context"], question=new_question, answer=new_answer)
else:
raise ValueError('No such mode')
def get_answer_type(dataset):
if dataset == 'strategyqa' or dataset == 'boolq':
return 'yes/no'
def get_verifier_label(verifier_response):
verifier_response_list = verifier_response.split('\n')
assert 'judgement: ' in verifier_response_list[-1].lower()
judgement = verifier_response_list[-1].split('judgement: ')[-1].strip()
if 'yes' in judgement.lower() and 'no' not in judgement.lower():
return 'pass'
else:
return 'fail'
def get_option(option_response):
option_response_list = option_response.split('\n')
assert 'option: ' in option_response_list[-1].lower()
option = option_response_list[-1].lower().split('option: ')[-1].strip()
return option
def generate_dataset(args, mode, task_description):
dataset = args.dataset
data_num = args.data_num
# initialize agents
generator = Agent(opt, hub, 'generator', mode, dataset, task_description)
verifier = Agent(opt, hub, 'verifier', mode, dataset, task_description)
option_generator = Agent(opt, hub, 'option_generator', mode, dataset, task_description)
# get test data
test_data, exist_instances, exist_ids = get_test_data(args, dataset, mode, data_num)
# generate dataset
for id in tqdm(range(data_num)):
each = test_data[id]
if id in exist_ids :
continue
new_inst = {'id': id, 'status': '', 'cause': '', 'context': '', 'question': '', 'answer': '', 'option': ''}
if each['status'] == 'fail':
new_inst = construct_new_inst(new_inst, status='fail', cause='Problem error: Original question failed')
exist_instances.append(new_inst)
output_result(dataset, mode, exist_instances)
continue
problem = get_problem(each, dataset, mode)
# 1. generate new instance
if args.debug:
print('---------------------------------------------------------------------')
print('Generate new instance:')
response = generator.ask(context=problem["context"], question=problem["question"], answer=problem["answer"], temp=0.0, max_tokens=512)
if mode not in ['paraphrase', 'addnoise']:
response_list = response.split('\n')
else:
response_list = [response]
if args.debug:
print(response)
new_inst = construct_new_inst_from_response(new_inst, problem, response_list, mode)
if new_inst['status'] == 'fail':
exist_instances.append(new_inst)
output_result(dataset, mode, exist_instances)
continue
if get_answer_type(dataset) == 'yes/no' and mode not in opt.sub_ability_mode:
if new_inst['answer'].lower() not in ['yes', 'no']:
new_inst = construct_new_inst(new_inst, status='fail', cause='Generator error: Answer type error')
exist_instances.append(new_inst)
output_result(dataset, mode, exist_instances)
continue
# 2. verify new instance
if args.debug:
print('---------------------------------------------------------------------')
print('Verify new instance:')
verifier_response = verifier.ask(context=new_inst["context"], question=new_inst["question"], answer=new_inst["answer"], temp=0.0, max_tokens=512)
if args.debug:
print(verifier_response)
label = get_verifier_label(verifier_response)
if label == 'fail':
new_inst = construct_new_inst(new_inst, status='fail', cause='Verify error: 1')
exist_instances.append(new_inst)
output_result(dataset, mode, exist_instances)
continue
else:
# 3. generate option
if args.debug:
print('---------------------------------------------------------------------')
print('Generate option:')
if "yes" == new_inst['answer'].lower():
if args.debug:
print('Directly generate option: No')
new_inst['option'] = "No"
elif "no" == new_inst['answer'].lower():
if args.debug:
print('Directly generate option: Yes')
new_inst['option'] = "Yes"
else:
option_response = option_generator.ask(context=new_inst["context"], question=new_inst["question"], answer=new_inst["answer"], temp=0.0, max_tokens=512)
if args.debug:
print(option_response)
option = get_option(option_response)
new_inst['option'] = option
# 4. verify option
if args.debug:
print('---------------------------------------------------------------------')
print('Verify option:')
verifier_response = verifier.ask(context=new_inst["context"], question=new_inst["question"], answer=new_inst["option"], temp=0.0, max_tokens=512)
if args.debug:
print(verifier_response)
label = get_verifier_label(verifier_response)
if label == 'pass':
new_inst = construct_new_inst(new_inst, status='fail', cause='Verify error: 2')
exist_instances.append(new_inst)
output_result(dataset, mode, exist_instances)
continue
else:
new_inst = construct_new_inst(new_inst, status='pass')
exist_instances.append(new_inst)
output_result(dataset, mode, exist_instances)
print(f'Length of generated dataset: {len([each for each in exist_instances if each["status"] == "pass"])}')
def main(args):
# set parameters
task_description = task_description_dict[args.dataset]
mode_list = [args.mode] if args.mode != 'all' else opt.rewrite_mode + opt.sub_ability_mode
if args.dataset not in opt.dataset_registry:
raise ValueError(f'Dataset:{args.dataset} not registered')
# generate dataset
for mode in mode_list:
if mode not in opt.rewrite_mode + opt.sub_ability_mode:
raise ValueError(f'Mode:{mode} not registered')
print(f'{args.dataset} evolve with mode {mode}...')
generate_dataset(args, mode, task_description)
if __name__ == '__main__':
args = get_args_parser().parse_args()
main(args)