-
Notifications
You must be signed in to change notification settings - Fork 0
/
ask_question.py
294 lines (232 loc) · 8.49 KB
/
ask_question.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
284
285
286
287
288
289
290
291
292
293
294
"""
Asks a question to a model.
Usage:
python ask_question.py --input_file <path to input file> --output_file <path to output file> --model <model name> --mode <baseline or few-shot>
Example:
python ask_question.py --input_file data/benchmark_samples/hendrycks_sample_0421.csv --output_file data/model_output/results_0421.csv --model text-davinci-003 --mode baseline
Authors: Peter Zhang and Isabella Borkovic
"""
import argparse
import datetime as dt
import string
from typing import List, Tuple
import backoff
import openai
import pandas as pd
TEST_FILE = f"data/benchmark_samples/hendrycks_sample_{dt.date.today().strftime('%m%d')}.csv"
MODES = ("baseline","cot","few-shot") # possible modes
# template paths
PROMPT_BASELINE_TEMPLATE = "templates/baseline_prompt_0217.txt"
FOLLOWUP_BASELINE_TEMPLATE = "templates/baseline_followup_0217.txt"
PROMPT_COT_TEMPLATE = "templates/cot_prompt_0218.txt"
FOLLOWUP_COT_TEMPLATE = "templates/cot_followup_0218.txt"
PROMPT_FEWSHOT_TEMPLATE = "templates/fewshot_prompt_0224.txt"
EXAMPLE_TEMPLATE = "templates/example_0224.txt"
PROMPT_IK_TEMPLATE = "templates/PIK_prompt.txt"
# tokens for reasoning
REASONING_LENGTH = 250
# choice cols in df
CHOICE_COLS = ("choice_A","choice_B","choice_C","choice_D")
openai.api_key = "" # your key here
def format_choices(choices: List[str]) -> str:
"""Adds lettering to a list of choices."""
assert len(choices) > 0
return "\n".join([f"({letter}) {choice}"
for letter, choice in zip(string.ascii_uppercase, choices)])
@backoff.on_exception(backoff.expo, [openai.error.RateLimitError, openai.error.APIError])
def completions_with_backoff(**kwargs):
"""Wrapper for openai.Completion.create() with backoff"""
return openai.Completion.create(**kwargs)
def letter_choice_completion(prompt: str, model: str):
"""Completion for a letter choice."""
completion = completions_with_backoff(
prompt=prompt,
temperature=0,
model=model,
max_tokens=1,
logprobs=5, # set higher because sometimes one of ABCD won"t be in the top 4
)
answer = completion["choices"][0]["text"]
logprobs = completion["choices"][0]["logprobs"]["top_logprobs"]
return answer, logprobs
def justification_completion(prompt: str, model: str):
"""Completion for a justification."""
completion = completions_with_backoff(
prompt=prompt,
temperature=0,
model=model,
max_tokens=REASONING_LENGTH
)
justification = completion["choices"][0]["text"]
return justification
def ask_question(
question: str,
choices: List[str],
model: str="text-davinci-003",
mode: str="baseline",
examples: str=None,
verbose: bool=False,
) -> Tuple[str, dict]:
"""
Quizzes the model on a multiple choice question.
Args:
question: the question to ask
choices: a list of choices
model: the model to use
mode: the mode to use
examples: examples to add to the prompt
Returns:
answer: the letter choice
logprobs: the logprobs of the letter choice
"""
assert mode in MODES
if mode=="baseline":
return ask_question_baseline(question=question, choices=choices, model=model, verbose=verbose)
if mode=="cot":
return ask_question_cot(question=question, choices=choices, model=model, verbose=verbose)
if mode=="few-shot":
return ask_question_fewshot(question=question, choices=choices, examples=examples, model=model, verbose=verbose)
def ask_question_baseline(
question: str,
choices: List[str],
model: str="text-davinci-003",
verbose: bool=False,
) -> Tuple[str, dict, str]:
"""Asks for a letter choice first and then the reasoning."""
# read prompt templates
prompt_template = open(PROMPT_BASELINE_TEMPLATE, "r").read()
followup_template = open(FOLLOWUP_BASELINE_TEMPLATE, "r").read()
# format letter choices
letter_choices = format_choices(choices)
# create prompt
prompt = prompt_template.format(
question=question,
letter_choices=letter_choices
)
# generate completion
answer, logprobs = letter_choice_completion(prompt, model)
# update prompt and generate completion
prompt += followup_template.format(answer=answer)
justification = justification_completion(prompt, model)
if verbose:
print(prompt)
return answer, logprobs, justification
def ask_question_cot(
question: str,
choices: List[str],
model: str="text-davinci-003",
verbose: bool=False,
) -> Tuple[str, dict, str]:
"""Asks for a reasoning first and then a letter choice."""
# read prompt templates
prompt_template = open(PROMPT_COT_TEMPLATE, "r").read()
followup_template = open(FOLLOWUP_COT_TEMPLATE, "r").read()
# format letter choices
letter_choices = format_choices(choices)
# create prompt
prompt = prompt_template.format(
question=question,
letter_choices=letter_choices,
)
# generate completion
justification = justification_completion(prompt, model)
# update prompt and generate completion
prompt += followup_template.format(justification=justification)
answer, logprobs = letter_choice_completion(prompt, model)
if verbose:
print(prompt)
return answer, logprobs, justification
def ask_question_fewshot(
question: str,
choices: List[str],
examples: str,
model: str="text-davinci-003",
verbose: bool=False,
) -> Tuple[str, dict, str]:
"""Ask question and adds examples."""
# read prompt templates
prompt_template = open(PROMPT_FEWSHOT_TEMPLATE, "r").read()
followup_template = open(FOLLOWUP_BASELINE_TEMPLATE, "r").read()
# format letter choices
letter_choices = format_choices(choices)
# create prompt
prompt = prompt_template.format(
examples=examples,
question=question,
letter_choices=letter_choices,
)
# generate completion
answer, logprobs = letter_choice_completion(prompt, model)
# update prompt and generate completion
prompt += followup_template.format(answer=answer)
justification = justification_completion(prompt, model)
if verbose:
print(prompt)
return answer, logprobs, justification
def format_example(
question: str,
choices: List[str],
correct_answer: str,
) -> str:
"""Formats an example for few-shot prompting."""
# read prompt templates
example_template = open(EXAMPLE_TEMPLATE, "r").read()
letter_choices = format_choices(choices)
return example_template.format(
question=question,
letter_choices=letter_choices,
answer=correct_answer,
)
def row_to_example(row: pd.Series) -> str:
"""Convert a row to an example."""
choices = [row[col] for col in CHOICE_COLS]
return format_example(
question=row["question"],
choices=choices,
correct_answer=row["correct_answer"],
)
def df_to_examples(df: pd.DataFrame, n_examples: int=None) -> str:
"""Convert a dataframe to examples."""
if n_examples:
df = df.iloc[:n_examples]
return "\n".join(df.apply(row_to_example, axis=1).tolist())
def ask_row(row, model: str="text-davinci-003", mode="baseline", examples=None, verbose=False):
"""Ask question but with a row."""
choices = [row[col] for col in CHOICE_COLS]
return ask_question(
question=row["question"],
choices=choices,
model=model,
mode=mode,
examples=examples,
verbose=verbose,
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--test_file", type=str, default=TEST_FILE)
parser.add_argument("--test_index", type=int, default=0)
parser.add_argument("--model", type=str, default="text-davinci-003")
parser.add_argument("--mode", type=str, default="baseline")
parser.add_argument("--num_examples", type=int, default=0)
args = parser.parse_args()
# load test data
df = pd.read_csv(args.test_file)
# convert to examples
if args.mode == "few-shot":
examples = df_to_examples(df, n_examples=args.num_examples)
else:
examples = None
# ask question
answer, logprobs, justification = ask_row(
row=df.iloc[args.test_index],
model=args.model,
mode=args.mode,
examples=examples,
verbose=True,
)
# print results
print(f"Answer: {answer}")
print(f"Logprobs: {logprobs}")
print(f"Justification: {justification}")
if __name__=="__main__":
main()