-
Notifications
You must be signed in to change notification settings - Fork 70
/
gpt_util.py
217 lines (180 loc) · 7.49 KB
/
gpt_util.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
import openai
import os
import numpy as np
import math
import codecs
from util.tokenizer import logit_mask
def normalize(probs):
return [float(i) / sum(probs) for i in probs]
def logprobs_to_probs(probs):
if isinstance(probs, list):
return [math.exp(x) for x in probs]
else:
return math.exp(probs)
def dict_logprobs_to_probs(prob_dict):
return {key: math.exp(prob_dict[key]) for key in prob_dict.keys()}
def total_logprob(response):
logprobs = response['logprobs']['token_logprobs']
logprobs = [i for i in logprobs if not math.isnan(i)]
return sum(logprobs)
def tokenize_ada(prompt):
response = openai.Completion.create(
engine='ada',
prompt=prompt,
max_tokens=0,
echo=True,
n=1,
logprobs=0
)
tokens = response.choices[0]["logprobs"]["tokens"]
positions = response.choices[0]["logprobs"]["text_offset"]
return tokens, positions
def prompt_probs(prompt, engine='ada'):
response = openai.Completion.create(
engine=engine,
prompt=prompt,
max_tokens=0,
echo=True,
n=1,
logprobs=0
)
positions = response.choices[0]["logprobs"]["text_offset"]
tokens = response.choices[0]["logprobs"]["tokens"]
logprobs = response.choices[0]["logprobs"]["token_logprobs"]
return logprobs, tokens, positions
# evaluates logL(prompt+target | prompt)
def conditional_logprob(prompt, target, engine='ada'):
combined = prompt + target
response = openai.Completion.create(
engine=engine,
prompt=combined,
max_tokens=0,
echo=True,
n=1,
logprobs=0
)
positions = response.choices[0]["logprobs"]["text_offset"]
logprobs = response.choices[0]["logprobs"]["token_logprobs"]
word_index = positions.index(len(prompt))
total_conditional_logprob = sum(logprobs[word_index:])
return total_conditional_logprob
# TODO use threading
# returns the conditional probabilities for each event happening after prompt
def event_probs(prompt, events, engine='ada'):
probs = []
for event in events:
logprob = conditional_logprob(prompt, event, engine)
probs.append(logprobs_to_probs(logprob))
normal_probs = normalize(probs)
return probs, normal_probs
# like event_probs, returns conditional probabilities (normalized & unnormalized) for each token occurring after prompt
def token_probs(prompt, tokens, engine='ada'):
pass
# returns a list of positions and counterfactual probability of token at position
# if token is not in top_logprobs, probability is treated as 0
# all positions if actual_token=None, else only positions where the actual token in response is actual_token
# TODO next sequence instead of next token
def counterfactual(response, token, actual_token=None, next_token=None, sort=True):
counterfactual_probs = []
tokens = response.choices[0]['logprobs']['tokens']
top_logprobs = response.choices[0]['logprobs']['top_logprobs']
positions = response.choices[0]['logprobs']['text_offset']
for i, probs in enumerate(top_logprobs):
if (actual_token is None and next_token is None) \
or actual_token == tokens[i] \
or (i < len(tokens) - 1 and next_token == tokens[i+1]):
if token in probs:
counterfactual_probs.append({'position': positions[i+1],
'prob': logprobs_to_probs(probs[token])})
else:
counterfactual_probs.append({'position': positions[i+1], 'prob': 0})
if sort:
counterfactual_probs = sorted(counterfactual_probs, key=lambda k: k['prob'])
return counterfactual_probs
# returns a list of substrings of content and
# logL(preprompt+substring+target | preprompt+substring) for each substring
def substring_probs(preprompt, content, target, engine='ada', quiet=0):
logprobs = []
substrings = []
_, positions = tokenize_ada(content)
for position in positions:
substring = content[:position]
prompt = preprompt + substring
logprob = conditional_logprob(prompt, target, engine)
logprobs.append(logprob)
substrings.append(substring)
if not quiet:
print(substring)
print('logprob: ', logprob)
return substrings, logprobs
# returns a list of substrings of content
# logL(substring+target | substring) for each substring
def token_conditional_logprob(content, target, engine='ada'):
response = openai.Completion.create(
engine=engine,
prompt=content,
max_tokens=0,
echo=True,
n=1,
logprobs=100
)
tokens = response.choices[0]['logprobs']['tokens']
top_logprobs = response.choices[0]['logprobs']['top_logprobs']
logprobs = []
substrings = []
substring = ''
for i, probs in enumerate(top_logprobs):
substrings.append(substring)
if target in probs:
logprobs.append(probs[target])
else:
logprobs.append(None)
substring += tokens[i]
return substrings, logprobs
def sort_logprobs(substrings, logprobs, n_top=None):
sorted_indices = np.argsort(logprobs)
top = []
if n_top is None:
n_top = len(sorted_indices)
for i in range(n_top):
top.append({'substring': substrings[sorted_indices[-(i + 1)]],
'logprob': logprobs[sorted_indices[-(i + 1)]]})
return top
def top_logprobs(preprompt, content, target, n_top=None, engine='ada', quiet=0):
substrings, logprobs = substring_probs(preprompt, content, target, engine, quiet)
return sort_logprobs(substrings, logprobs, n_top)
def decibels(prior, evidence, target, engine='ada'):
prior_target_logprob = conditional_logprob(prompt=prior, target=target, engine=engine)
evidence_target_logprob = conditional_logprob(prompt=evidence, target=target, engine=engine)
return (evidence_target_logprob - prior_target_logprob), prior_target_logprob, evidence_target_logprob
def parse_stop(stop_string):
return codecs.decode(stop_string, "unicode-escape").split('|')
def parse_logit_bias(logit_string):
biases = codecs.decode(logit_string, "unicode-escape").split('|')
bias_dict = {}
for b in biases:
bias_parts = b.split(':')
token = bias_parts[0]
bias = int(bias_parts[1])
bias_dict[token] = bias
return logit_mask(bias_dict)
def get_correct_key(model_type, kwargs={}):
if model_type == 'gooseai':
# openai.api_base = openai.api_base if openai.api_base else "https://api.goose.ai/v1"
gooseai_api_key = kwargs.get('GOOSEAI_API_KEY', None)
api_key = gooseai_api_key if gooseai_api_key else os.environ.get("GOOSEAI_API_KEY", None)
organization = None
if model_type == 'together':
togetherai_api_key = kwargs.get('TOGETHERAI_API_KEY', None)
api_key = togetherai_api_key if togetherai_api_key else os.environ.get("TOGETHERAI_API_KEY", None)
organization = None
elif model_type in ('openai', 'openai-custom', 'openai-chat'):
# openai.api_base = openai.api_base if openai.api_base else "https://api.openai.com/v1"
openai_api_key = kwargs.get('OPENAI_API_KEY', None)
api_key = openai_api_key if openai_api_key else os.environ.get("OPENAI_API_KEY", None)
openai_organization = kwargs.get('OPENAI_ORGANIZATION', None)
organization = openai_organization if openai_organization else os.environ.get("OPENAI_ORGANIZATION", None)
else:
api_key = None
organization = None
return api_key, organization