-
Notifications
You must be signed in to change notification settings - Fork 3
/
inference.py
179 lines (153 loc) · 6.79 KB
/
inference.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
import os
from tqdm import tqdm
import json
import os
import openai
from tqdm import tqdm
import argparse
import multiprocessing
from copy import deepcopy
from functools import partial
prompt_library = {
"MCQ": "In this problem, only one option will be correct. Give a detailed solution and end the solution with the final answer.",
"MCQ(multiple)": "In this problem, multiple options can be correct. Give a detailed solution and end the solution with the final answer.",
"Integer": "In this problem, the final answer will be a non-negative integer. Give a detailed solution and end the solution with the final answer.",
"Numeric": "In this problem, the final will be a numeric value. Give the numerical answer correct upto the 2nd decimal digit. Give a detailed solution and end the solution with the final answer.",
}
few_shot_examples = json.load(open('data/few_shot_examples.json'))
def write_in_file(response_file, response_dict, question, mode, model_nickname):
if os.path.exists(response_file):
with open(response_file, 'r') as infile:
responses = json.load(infile)
else:
responses = []
found = False
for i, old_resp in enumerate(responses):
if old_resp['description'] == question['description'] and old_resp['index'] == question['index']:
responses[i][f"{model_nickname}_{mode}_response" ] = response_dict[f"{model_nickname}_{mode}_response"]
found = True
break
if not found:
responses.append(response_dict)
json.dump(sorted(responses, key=lambda elem: (elem['description'], elem['index'])), open(response_file, 'w'), indent=4)
print(f"####UPDATED {response_file}, Current size : {len(responses)}####")
def get_response(question,model, model_nickname, mode, response_file, lock):
response_dict = deepcopy(question)
prefix_prompt = prompt_library[question['type']]
suffix_prompt = ""
if mode in ['CoT', 'CoT+SC', 'CoT+Exam'] :
suffix_prompt = "Let's think step by step.\n"
ques = question["question"]
stripped_ques = ques.replace("\n\n", "\n").strip()
if mode in ['CoT+OneShot', 'CoT', 'CoT+SC', 'CoT+Exam']:
if mode == 'CoT+Exam':
if response_dict['type'] in ['MCQ', 'MCQ(multiple)']:
if response_dict['type'] == 'MCQ':
exam_prompt = "If the answer is wrong, you'll be given -1 marks. If the answer is correct, you'll be given +3 marks. If you're unsure of the answer, you can skip the question, and you'll be given 0 marks."
else:
exam_prompt = "If any of the options in the final answer is wrong, you'll be given -2 marks. If all the options are correct, you'll be given +4 marks. If some of the options are correct, you'll be given +1 for each correct option. If you're unsure of the answer, you can skip the question, and you'll be given 0 marks."
prompt = prefix_prompt + " " + exam_prompt + "\n\n" + "Problem: " + stripped_ques + "\nSolution: " + suffix_prompt
else:
print("No point doing this for Numeric/Integer questions since there is no negative marking...")
breakpoint()
else:
if mode == 'CoT+OneShot':
ex = few_shot_examples[question['subject']][question['type']]
prompt = prefix_prompt + "\n\n" + "Problem: " + ex['problem'] + "\nSolution: " + ex['solution'] + "\n\n" + "Problem: " + stripped_ques + "\nSolution: "
else:
prompt = prefix_prompt + "\n\n" + "Problem: " + stripped_ques + "\nSolution: " + suffix_prompt
else:
prompt = prefix_prompt + "\n\n" + "Problem: " + stripped_ques + suffix_prompt
prompt = prompt.strip()
response_dict[f"prompt"] = prompt
num_retries = 0
print(f'Question: {question["description"]}, Index: {question["index"]}, Model: {model_nickname}, Mode: {mode}, query begins')
while True:
try:
if model in ["text-davinci-003", "text-davinci-002", 'davinci-002']:
response = openai.Completion.create(
model=model,
prompt=prompt,
max_tokens=2048,
temperature=0 if mode in ['CoT', 'normal', 'CoT+Exam'] else 0.5,
n=1 if mode in ['CoT', 'normal', 'CoT+Exam'] else 3
)
else:
response = openai.ChatCompletion.create(
model=model,
messages=[
{"role": "system", "content": ""},
{"role": "user", "content": prompt}
],
max_tokens=2048,
temperature=0 if mode in ['CoT+OneShot', 'CoT', 'normal', 'CoT+Exam'] else 0.5,
n=1 if mode in ['CoT+OneShot', 'CoT', 'normal', 'CoT+Exam'] else 8
)
lock.acquire()
response_dict[f"{model_nickname}_{mode}_response"] = response
write_in_file(response_file, response_dict, question, mode, model_nickname)
lock.release()
break
except Exception as e:
num_retries += 1
print("Failure!", e)
return
def main():
'''
The code can restart from the already done questions in case there is a failure midpoint.
'''
args = argparse.ArgumentParser()
args.add_argument('--model', default='gpt-3.5-turbo')
args.add_argument('--data', default='data/dataset.json')
args.add_argument('--mode', default='normal')
args.add_argument('--num_procs', default=1, type=int)
args.add_argument('--max_questions', default=1, type=int)
args = args.parse_args()
openai.organization = os.getenv("OPENAI_ORG")
openai.api_key = os.getenv("OPENAI_API_KEY")
model_nickname = {
"davinci-002": "davinci-002",
"text-davinci-003": "GPT3",
"gpt-3.5-turbo": "GPT3.5",
"gpt-4-0613": "GPT4_0613",
"gpt-4-0314": "GPT4"
}
assert args.model in model_nickname.keys()
assert args.mode in ['normal', 'CoT', 'CoT+OneShot', 'CoT+Exam', 'CoT+SC']
out_file_dir = f'responses/{model_nickname[args.model]}_{args.mode}_responses'
out_file = os.path.join(out_file_dir, 'responses.json')
questions = json.load(open(args.data))
rem_ques = []
if os.path.exists(out_file):
for question in tqdm(questions[:args.max_questions]):
if os.path.exists(out_file):
with open(out_file, 'r') as infile:
responses = json.load(infile)
found = False
for i, old_resp in enumerate(responses):
if question['type'] in ['Numeric', 'Integer'] and args.mode == 'CoT+Exam':
found = True
if old_resp['description'] == question['description'] and old_resp['index'] == question['index']:
found = all([old_resp.get(
f"{model_nickname[args.model]}_{args.mode}_response", False) for model in [args.model]])
if found:
print("This question has already been done")
else:
rem_ques.append(question)
else:
os.makedirs(out_file_dir, exist_ok=True)
if args.mode == 'CoT+Exam':
rem_ques = []
for q in questions:
if q['type'] in ['MCQ', 'MCQ(multiple)']:
rem_ques.append(q)
else:
rem_ques = questions[:args.max_questions]
print(f"There are {len(rem_ques)} problems remaining")
manager = multiprocessing.Manager()
lock = manager.Lock()
pool = multiprocessing.Pool(args.num_procs)
f = partial(get_response, model=args.model, model_nickname=model_nickname[args.model], mode=args.mode, response_file=out_file, lock=lock)
pool.map(f, rem_ques)
if __name__ == '__main__':
main()