-
Notifications
You must be signed in to change notification settings - Fork 58
/
conversers.py
240 lines (202 loc) · 9.26 KB
/
conversers.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
import common
from language_models import GPT, Claude, PaLM, HuggingFace
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from config import VICUNA_PATH, LLAMA_PATH, ATTACK_TEMP, TARGET_TEMP, ATTACK_TOP_P, TARGET_TOP_P
def load_attack_and_target_models(args):
# Load attack model and tokenizer
attackLM = AttackLM(model_name = args.attack_model,
max_n_tokens = args.attack_max_n_tokens,
max_n_attack_attempts = args.max_n_attack_attempts,
temperature = ATTACK_TEMP, # init to 1
top_p = ATTACK_TOP_P, # init to 0.9
)
preloaded_model = None
if args.attack_model == args.target_model:
print("Using same attack and target model. Using previously loaded model.")
preloaded_model = attackLM.model
targetLM = TargetLM(model_name = args.target_model,
max_n_tokens = args.target_max_n_tokens,
temperature = TARGET_TEMP, # init to 0
top_p = TARGET_TOP_P, # init to 1
preloaded_model = preloaded_model,
)
return attackLM, targetLM
class AttackLM():
"""
Base class for attacker language models.
Generates attacks for conversations using a language model. The self.model attribute contains the underlying generation model.
"""
def __init__(self,
model_name: str,
max_n_tokens: int,
max_n_attack_attempts: int,
temperature: float,
top_p: float):
self.model_name = model_name
self.temperature = temperature
self.max_n_tokens = max_n_tokens
self.max_n_attack_attempts = max_n_attack_attempts
self.top_p = top_p
self.model, self.template = load_indiv_model(model_name)
if "vicuna" in model_name or "llama" in model_name:
self.model.extend_eos_tokens()
def get_attack(self, convs_list, prompts_list):
"""
Generates responses for a batch of conversations and prompts using a language model.
Only valid outputs in proper JSON format are returned. If an output isn't generated
successfully after max_n_attack_attempts, it's returned as None.
Parameters:
- convs_list: List of conversation objects.
- prompts_list: List of prompts corresponding to each conversation.
Returns:
- List of generated outputs (dictionaries) or None for failed generations.
"""
assert len(convs_list) == len(prompts_list), "Mismatch between number of conversations and prompts."
batchsize = len(convs_list)
indices_to_regenerate = list(range(batchsize))
valid_outputs = [None] * batchsize
# Initalize the attack model's generated output to match format
if len(convs_list[0].messages) == 0:
init_message = """{\"improvement\": \"\",\"prompt\": \""""
else:
init_message = """{\"improvement\": \""""
full_prompts = []
# Add prompts and initial seeding messages to conversations (only once)
for conv, prompt in zip(convs_list, prompts_list):
conv.append_message(conv.roles[0], prompt)
# Get prompts
if "gpt" in self.model_name:
full_prompts.append(conv.to_openai_api_messages())
else:
conv.append_message(conv.roles[1], init_message)
full_prompts.append(conv.get_prompt()[:-len(conv.sep2)])
for attempt in range(self.max_n_attack_attempts):
# Subset conversations based on indices to regenerate
full_prompts_subset = [full_prompts[i] for i in indices_to_regenerate]
# Generate outputs
outputs_list = self.model.batched_generate(full_prompts_subset,
max_n_tokens = self.max_n_tokens,
temperature = self.temperature,
top_p = self.top_p
)
# Check for valid outputs and update the list
new_indices_to_regenerate = []
for i, full_output in enumerate(outputs_list):
orig_index = indices_to_regenerate[i]
if "gpt" not in self.model_name:
full_output = init_message + full_output
attack_dict, json_str = common.extract_json(full_output)
if attack_dict is not None:
valid_outputs[orig_index] = attack_dict
convs_list[orig_index].update_last_message(json_str) # Update the conversation with valid generation
else:
new_indices_to_regenerate.append(orig_index)
# Update indices to regenerate for the next iteration
indices_to_regenerate = new_indices_to_regenerate
# If all outputs are valid, break
if not indices_to_regenerate:
break
if any([output for output in valid_outputs if output is None]):
print(f"Failed to generate output after {self.max_n_attack_attempts} attempts. Terminating.")
return valid_outputs
class TargetLM():
"""
Base class for target language models.
Generates responses for prompts using a language model. The self.model attribute contains the underlying generation model.
"""
def __init__(self,
model_name: str,
max_n_tokens: int,
temperature: float,
top_p: float,
preloaded_model: object = None):
self.model_name = model_name
self.temperature = temperature
self.max_n_tokens = max_n_tokens
self.top_p = top_p
if preloaded_model is None:
self.model, self.template = load_indiv_model(model_name)
else:
self.model = preloaded_model
_, self.template = get_model_path_and_template(model_name)
def get_response(self, prompts_list):
batchsize = len(prompts_list)
convs_list = [common.conv_template(self.template) for _ in range(batchsize)]
full_prompts = []
for conv, prompt in zip(convs_list, prompts_list):
conv.append_message(conv.roles[0], prompt)
if "gpt" in self.model_name:
# Openai does not have separators
full_prompts.append(conv.to_openai_api_messages())
elif "palm" in self.model_name:
full_prompts.append(conv.messages[-1][1])
else:
conv.append_message(conv.roles[1], None)
full_prompts.append(conv.get_prompt())
outputs_list = self.model.batched_generate(full_prompts,
max_n_tokens = self.max_n_tokens,
temperature = self.temperature,
top_p = self.top_p
)
return outputs_list
def load_indiv_model(model_name, device=None):
model_path, template = get_model_path_and_template(model_name)
if model_name in ["gpt-3.5-turbo", "gpt-4"]:
lm = GPT(model_name)
elif model_name in ["claude-2", "claude-instant-1"]:
lm = Claude(model_name)
elif model_name in ["palm-2"]:
lm = PaLM(model_name)
else:
model = AutoModelForCausalLM.from_pretrained(
model_path,
torch_dtype=torch.float16,
low_cpu_mem_usage=True,device_map="auto").eval()
tokenizer = AutoTokenizer.from_pretrained(
model_path,
use_fast=False
)
if 'llama-2' in model_path.lower():
tokenizer.pad_token = tokenizer.unk_token
tokenizer.padding_side = 'left'
if 'vicuna' in model_path.lower():
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = 'left'
if not tokenizer.pad_token:
tokenizer.pad_token = tokenizer.eos_token
lm = HuggingFace(model_name, model, tokenizer)
return lm, template
def get_model_path_and_template(model_name):
full_model_dict={
"gpt-4":{
"path":"gpt-4",
"template":"gpt-4"
},
"gpt-3.5-turbo": {
"path":"gpt-3.5-turbo",
"template":"gpt-3.5-turbo"
},
"vicuna":{
"path":VICUNA_PATH,
"template":"vicuna_v1.1"
},
"llama-2":{
"path":LLAMA_PATH,
"template":"llama-2"
},
"claude-instant-1":{
"path":"claude-instant-1",
"template":"claude-instant-1"
},
"claude-2":{
"path":"claude-2",
"template":"claude-2"
},
"palm-2":{
"path":"palm-2",
"template":"palm-2"
}
}
path, template = full_model_dict[model_name]["path"], full_model_dict[model_name]["template"]
return path, template