-
Notifications
You must be signed in to change notification settings - Fork 0
/
reflexion.py
253 lines (231 loc) · 11.6 KB
/
reflexion.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
import tqdm
from utils import enumerate_resume, make_printv, write_jsonl, resume_success_count
from generators import generator_factory, model_factory
import ezsheets
import time
# from langchain_community.document_loaders.csv_loader import CSVLoader
# from langchain_community.document_loaders import JSONLoader
# from langchain_chroma import Chroma
# from langchain_community.embeddings import HuggingFaceEmbeddings
# from langchain.storage import LocalFileStore
# from langchain.embeddings import CacheBackedEmbeddings
from typing import List
def run_reflexion(
dataset: List[dict],
pe_model_name: str,
ue_model_name: str,
act_model_name: str,
parser_model_name: str,
language: str,
max_iters: int,
pass_at_k: int,
log_path: str,
verbose: bool,
mem_len: int,
p_threshold: int,
is_leetcode: bool = False,
no_utility: bool = False,
cot: bool = False,
rag_data_path: str = '',
rag_num: int = 5,
rag_embed_cache_dir: str = '',
) -> None:
gen = generator_factory(language)
pe_model = model_factory(pe_model_name)
ue_model = model_factory(ue_model_name)
act_model = model_factory(act_model_name)
parser_model = model_factory(parser_model_name)
# if rag_data_path.endswith('.csv'):
# loader = CSVLoader(file_path=rag_data_path, source_column="text")
# # "./programming_runs/benchmarks/Wiki_People/DBP_wiki_data.csv"
# else:
# assert rag_data_path.endswith('.jsonl')
#
# def metadata_func(record: dict, metadata: dict) -> dict:
# metadata['l1'] = record.get("l1")
# metadata['l2'] = record.get("l2")
# metadata['l3'] = record.get('l3')
# name = record.get('wiki_name')
# name = name.replace('_', ' ')
# if '(' in name:
# name = name.replace(name[name.index('('):name.index(')') + 1], '')
# metadata['wiki_name'] = name
# metadata['word_count'] = record.get('word_count')
# return metadata
#
# loader = JSONLoader(
# file_path=rag_data_path,
# # './programming_runs/benchmarks/Wiki_People/All_data_for_retrieval.jsonl'
# jq_schema='.',
# content_key='text',
# metadata_func=metadata_func,
# json_lines=True)
# data = loader.load()
# embeddings = HuggingFaceEmbeddings(model_name="BAAI/llm-embedder")
# # "all-MiniLM-L6-v2"
# store = LocalFileStore(rag_embed_cache_dir)
# cached_embedder = CacheBackedEmbeddings.from_bytes_store(
# embeddings, store, namespace="llm-embedder"
# )
# vectorstore = Chroma.from_documents(documents=data, embedding=cached_embedder)
# retriever = vectorstore.as_retriever(search_type="similarity", search_kwargs={"k": rag_num})
for i, item in enumerate_resume(tqdm.tqdm(dataset[107 + 39 + 15:]), log_path):
# try:
cur_pass = 0
complete = False
acc_reward = 0
privacy_reflections = []
utility_reflections = []
rewritings = []
people = item["people"] if language == 'wiki' else {item['feature']: item['personality'][item['feature']]}
if item['label'] == 'Medician':
item['label'] = 'Physician'
if cot:
detection_i = gen.detect(item["text"] if language == 'wiki' else item['response'].replace('\n', ''), act_model)
while cur_pass < pass_at_k and not complete:
privacy_reflections.append(f"pass: {cur_pass}")
utility_reflections.append(f"pass: {cur_pass}")
rewritings.append(f"pass: {cur_pass}")
# first attempt
cur_rewriting = gen.rewrite(item["text"] if language == 'wiki' else item['response'].replace('\n', ''), item['label'], people, act_model, parser_model, "simple", cot=cot,
detection_result=detection_i['raw_response'] if cot else None,
temperature=0.0)
rewritings.append(cur_rewriting)
privacy_evaluation = gen.privacy_reflex(pe_model, rewritings[-1]['Anonymized text'], people, p_threshold,
no_utility, None)
privacy_score = privacy_evaluation["Confirmation"]
privacy_feedback = privacy_evaluation["Advice"]
privacy_reflections.append(privacy_evaluation)
if not no_utility:
utility_evaluation = gen.utility_reflex(item['text'] if language == 'wiki' else item['response'].replace('\n', ''),
ue_model, rewritings[-1]['Anonymized text'],
item['label'] if language == 'wiki' else item['personality']['occupation'],
privacy_score)
utility_score = utility_evaluation["Confirmation"]
utility_feedback = utility_evaluation["Advice"]
utility_reflections.append(utility_evaluation)
else:
utility_evaluation = {'Confirmation': 'Yes', 'Advice': ''}
utility_score = utility_evaluation["Confirmation"]
utility_feedback = utility_evaluation["Advice"]
utility_reflections.append(utility_evaluation)
# if solved, exit early
if privacy_score == 'No' and utility_score == 'Yes':
complete = True
acc_reward = p_threshold + 1 + 100 if not no_utility else p_threshold + 1
break
cur_iter = 1
complete = False
acc_reward = 0
while cur_iter <= max_iters:
# apply self-reflection in the next attempt
if no_utility:
prev_rewriting = cur_rewriting['raw_response']
acc_reward += int(privacy_evaluation['rank'])
else:
prev_rewriting = ''
h_idx = 1
acc_reward = 0
if len(rewritings) > mem_len:
p_rer = rewritings[-mem_len:]
p_pr = privacy_reflections[-mem_len:]
p_ur = utility_reflections[-mem_len:]
else:
p_rer = rewritings
p_pr = privacy_reflections
p_ur = utility_reflections
for rewriting, p_r, u_r in zip(p_rer, p_pr, p_ur):
if type(rewriting) is str:
continue
prev_rewriting += f"Edition: {h_idx}\nEditing results; {rewriting['Anonymized text']}\nPrivacy score: {p_r['rank']}\nUtility score: {u_r['Confidence Score']}\n"
if p_r['Confirmation'] == 'Yes':
prev_rewriting += f"Reward: {p_r['rank']}\n\n"
acc_reward += int(p_r['rank'])
else:
prev_rewriting += f"Reward: {u_r['Confidence Score']}\n\n"
acc_reward += int(u_r['Confidence Score'])
h_idx = h_idx + 1
cur_rewriting = gen.rewrite(
input_text=item["text"] if language == 'wiki' else item['response'].replace('\n', ''),
label=item['label'],
people=people,
act_model=act_model,
parser_model=parser_model,
cot=cot,
strategy="reflexion",
prev_rewriting=prev_rewriting,
reflection_privacy=privacy_feedback,
reflection_utility=utility_feedback,
privacy_score=privacy_score,
utility_score=utility_score,
detection_result=None,
p_threshold=p_threshold if language == 'wiki' else 7,
no_utility=no_utility
)
rewritings.append(cur_rewriting)
# get self-reflection
text_tobe_evaluated = cur_rewriting['Anonymized text']
privacy_evaluation = gen.privacy_reflex(pe_model, text_tobe_evaluated, people, p_threshold, no_utility,
None)
privacy_score = privacy_evaluation["Confirmation"]
privacy_feedback = privacy_evaluation["Advice"]
privacy_reflections.append(privacy_evaluation)
if not no_utility:
utility_evaluation = gen.utility_reflex(item['text'] if language == 'wiki' else item['response'].replace('\n', ''),
ue_model, text_tobe_evaluated,
item['label'] if language == 'wiki' else item['personality']['occupation'],
privacy_score)
utility_score = utility_evaluation["Confirmation"]
utility_feedback = utility_evaluation["Advice"]
utility_reflections.append(utility_evaluation)
else:
utility_evaluation = {'Confirmation': 'Yes', 'Advice': ''}
utility_score = utility_evaluation["Confirmation"]
utility_feedback = utility_evaluation["Advice"]
utility_reflections.append(utility_evaluation)
# if solved, check if it passes the real tests, exit early
if privacy_score == 'No' and utility_score == 'Yes':
complete = True
break
cur_iter += 1
cur_pass += 1
item["rewritings"] = rewritings
item["privacy_reflections"] = privacy_reflections
item["utility_reflections"] = utility_reflections
item["complete"] = 'False' if not complete else 'True'
item["acc_reward"] = acc_reward
if cot:
item["detection_result"] = detection_i
write_jsonl(log_path, [item], append=True)
act_model.print_usage()
pe_model.print_usage()
ue_model.print_usage()
parser_model.print_usage()
print(f"log path: {log_path}\n")
# except Exception as e:
# act_model.print_usage()
# pe_model.print_usage()
# ue_model.print_usage()
# parser_model.print_usage()
# write_jsonl(log_path, [{'status': 'Failed'}], append=True)
# print(f"{e}\n{i}-th example failed")
ss = ezsheets.Spreadsheet('1-uHO5DnE32WmImaucvHaVMvasO2mGh2eqWfWYksXljI')
sheet = ss[0]
update_idx = sheet.getColumn(1).index('') + 1
update_row = sheet.getRow(update_idx)
name2column = {'gpt-35-turbo-0301': 7, 'gpt-4': 1, 'gpt4-turbo-128k': 4, 'gpt-4-turbo-preview': 10}
name2prompt_tokens = {'gpt-35-turbo-0301': 0, 'gpt-4': 0, 'gpt4-turbo-128k': 0, 'gpt-4-turbo-preview': 0}
name2completion_tokens = {'gpt-35-turbo-0301': 0, 'gpt-4': 0, 'gpt4-turbo-128k': 0, 'gpt-4-turbo-preview': 0}
model_list = [act_model, pe_model, ue_model, parser_model]
for model in model_list:
if model.name in name2column.keys():
name2prompt_tokens[model.name] += model.prompt_tokens
name2completion_tokens[model.name] += model.completion_tokens
for k, v in name2prompt_tokens.items():
update_row[name2column[k]] = v
for k, v in name2completion_tokens.items():
update_row[name2column[k] + 1] = v
update_row[0] = time.ctime()
sheet.refresh()
update_idx = sheet.getColumn(1).index('') + 1
sheet.updateRow(update_idx, update_row)