Spaces:
Running
on
Zero
Running
on
Zero
File size: 10,984 Bytes
85e172b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 |
import os
import copy
import numpy as np
from tqdm import tqdm
import torch
from . import compute_wb
from util import extraction
from util import utils
def extract_subject_feature(
request,
model,
tok,
layer,
module_template
):
""" Extracts the subject feature from a model for a single request, whole prompt for attacks
"""
# retrieves the last token representation of `subject`
feature_vector = extraction.extract_features_at_tokens(
model,
tok,
prompts = [request["prompt"]],
subjects = [request["subject"]],
layer = layer,
module_template = module_template,
tok_type = 'subject_final',
track = 'in',
batch_size = 1,
return_logits = False,
verbose = False
)[0]
return feature_vector
def augment_prompt(prompt, aug_mode, num_aug, stopwords=['{}'], size_limit=10):
""" Use of nlpaug to augment texutal prompts
"""
import nlpaug.augmenter.char as nac
import nlpaug.augmenter.word as naw
if aug_mode == 'KeyboardAug':
aug = nac.KeyboardAug(stopwords=stopwords, aug_char_max=size_limit)
augmented_prompts = aug.augment(prompt, n=num_aug)
elif aug_mode == 'OcrAug':
aug = nac.OcrAug(stopwords=stopwords, aug_char_max=size_limit)
augmented_prompts = aug.augment(prompt, n=num_aug)
elif aug_mode == 'RandomCharInsert':
aug = nac.RandomCharAug(stopwords=stopwords, action="insert", aug_char_max=size_limit)
augmented_prompts = aug.augment(prompt, n=num_aug)
elif aug_mode == 'RandomCharSubstitute':
aug = nac.RandomCharAug(stopwords=stopwords, action="substitute", aug_char_max=size_limit)
augmented_prompts = aug.augment(prompt, n=num_aug)
elif aug_mode == 'SpellingAug':
aug = naw.SpellingAug(stopwords=stopwords, aug_max=size_limit)
augmented_prompts = aug.augment(prompt, n=num_aug)
else:
raise AssertionError('Augmentation mode not supported: {}'.format(aug_mode))
return augmented_prompts
def iterative_augment_prompt(
aug_portion,
aug_mode='KeyboardAug',
size_limit = 1,
same_length = False,
num_aug = 10000
):
""" Iterative augmentation until size limit reqched
"""
all_augmented_prompts = []
count = 0
portion_length = len(aug_portion)
while True:
augmented_prompts = augment_prompt(aug_portion, aug_mode, num_aug, size_limit=size_limit)
if aug_portion.endswith(' '):
augmented_prompts = [augmented_prompt + ' ' for augmented_prompt in augmented_prompts]
all_augmented_prompts = all_augmented_prompts + augmented_prompts
# find unique preprompts
unique_augmented_prompts = np.unique(all_augmented_prompts)
# same length
if same_length:
lengths = np.array([len(t) for t in unique_augmented_prompts])
unique_augmented_prompts = unique_augmented_prompts[lengths == portion_length]
if (len(unique_augmented_prompts) >= num_aug) or (count > 30):
break
count += 1
augmented_prompts = unique_augmented_prompts[:num_aug]
return augmented_prompts
def extract_augmentations(
model,
tok,
request,
layers,
module_template = 'transformer.h.{}.mlp.c_fc',
tok_type = 'last',
num_aug = 2000,
aug_mode = 'KeyboardAug',
size_limit = 1,
batch_size = 64,
aug_portion = 'prompt',
static_context = None,
return_logits = True,
augmented_cache = None,
return_augs_only = False,
include_original = True,
include_comparaitve = False,
return_features = True,
verbose = False
):
""" Make text augmentations and extract features
"""
if type(layers) == int: layers = [layers]
# find prompt and subject of request
word = request["subject"]
prompt = request["prompt"]
# find portion of text to augment
if aug_portion == 'context':
pre_prompt = ''
to_aug_text = static_context
post_prompt = prompt
same_length = False
elif aug_portion == 'wikipedia':
pre_prompt = ''
to_aug_text = None
post_prompt = prompt
same_length = False
elif aug_portion == 'prompt':
to_aug_text = prompt.format(word)
same_length = False
else:
raise ValueError('invalid option for which portion to augment:', aug_portion)
# perform text augmentation
if augmented_cache is not None:
if type(augmented_cache)==str:
augmented_cache = utils.loadjson(augmented_cache)['augmented_cache']
if len(augmented_cache)> num_aug:
augmented_cache = np.random.choice(augmented_cache, num_aug, replace=False)
augmented_texts = augmented_cache
else:
augmented_texts = iterative_augment_prompt(
aug_portion=to_aug_text,
aug_mode=aug_mode,
size_limit=size_limit,
same_length = same_length,
num_aug = num_aug
)
if return_augs_only:
return augmented_texts
# add original as first one
if (to_aug_text is not None) and (include_original):
augmented_texts = np.array([to_aug_text] + list(augmented_texts))
else:
augmented_texts = np.array(augmented_texts)
# process back to subject and prompts
if aug_portion in ['context']:
aug_prompts = [pre_prompt + a + post_prompt for a in augmented_texts]
aug_subjects = [word for a in augmented_texts]
if include_comparaitve:
aug_prompts.append(static_context+'{}')
aug_subjects.append('')
aug_prompts.append(prompt)
aug_subjects.append(word)
elif aug_portion == 'wikipedia':
aug_prompts = [pre_prompt + a + post_prompt for a in augmented_texts]
aug_subjects = [word for a in augmented_texts]
if include_comparaitve:
aug_prompts.append(augmented_texts[0]+'{}')
aug_subjects.append('')
aug_prompts = [prompt] + aug_prompts
aug_subjects = [word] + aug_subjects
elif aug_portion == 'prompt':
# use same subject text indices since we have static length augmentation
start_idx = len(prompt.split('{}')[0])
end_idx = len(prompt.split('{}')[0]) + len(word)
aug_subjects = augmented_texts
aug_prompts = ['{}' for i in range(len(augmented_texts))]
if return_features:
# extract feature cloud
layer_names = [module_template.format(l) for l in layers]
extraction_return = extraction.extract_multilayer_at_tokens(
model,
tok,
prompts = aug_prompts,
subjects = aug_subjects,
layers = layer_names,
module_template = None,
tok_type = tok_type,
track = 'in',
return_logits = return_logits,
batch_size = batch_size,
verbose = verbose
)
feature_cloud = torch.stack([extraction_return[l]['in'] for l in layer_names])
else:
feature_cloud = None
aug_prompts = np.array(aug_prompts)
aug_subjects = np.array(aug_subjects)
if return_logits:
aug_logits = extraction_return['tok_predictions']
else:
aug_logits = None
return aug_prompts, aug_subjects, feature_cloud, aug_logits
def convert_to_prompt_only_request(request):
new_request = copy.deepcopy(request)
new_request['prompt'] = '{}'
new_request['subject'] = request['prompt'].format(request['subject'])
return new_request
def extract_target(
request,
model,
tok,
layer,
hparams,
mode = 'prompt'
):
""" Function to extract target features
"""
target_set = {}
if mode in ['prompt', 'origin_prompt', 'origin_context', 'origin_wikipedia']:
if (mode == 'prompt') and (request['prompt'] != '{}'):
raise ValueError('Mode [prompt] only works for empty request prompt [{}]')
# find w1 input of target prompt
target_set['w1_input'] = extract_subject_feature(
request,
model,
tok,
layer = layer,
module_template = hparams['rewrite_module_tmp'],
)
target_set['Y_current'] = np.array(
target_set['w1_input'].cpu().numpy(), dtype=np.float32)
elif mode in ['context', 'instructions']:
# find w1 input of just context
target_set['ctx_request'] = copy.deepcopy(request)
target_set['ctx_request']['prompt'] = "{}"
target_set['ctx_request']['subject'] = hparams['static_context']
target_set['w1_context'] = extract_subject_feature(
target_set['ctx_request'],
model,
tok,
layer = layer,
module_template = hparams['rewrite_module_tmp'],
)
target_set['Y_context'] = np.array(
target_set['w1_context'].cpu().numpy(), dtype=np.float32)
# find w1 input of original request
target_set['org_request'] = copy.deepcopy(request)
target_set['org_request']['prompt'] = target_set['org_request']['prompt'].split(hparams['static_context'])[-1]
target_set['org_request'] = convert_to_prompt_only_request(target_set['org_request'])
# find w1 input of target subject NOTE: need to change load from fcloud to subject pickle file
target_set['w1_org'] = extract_subject_feature(
target_set['org_request'],
model,
tok,
layer = layer,
module_template = hparams['rewrite_module_tmp'],
)
target_set['Y_org_current'] = np.array(
target_set['w1_org'].cpu().numpy(), dtype=np.float32)
# find w1 input of static context + original request
target_set['oap_request'] = copy.deepcopy(request)
if not target_set['oap_request']['prompt'].startswith(hparams['static_context']):
target_set['oap_request']['prompt'] = hparams['static_context'] + target_set['oap_request']['prompt']
target_set['oap_request'] = convert_to_prompt_only_request(target_set['oap_request'])
target_set['w1_oap'] = extract_subject_feature(
target_set['oap_request'],
model,
tok,
layer = layer,
module_template = hparams['rewrite_module_tmp'],
)
target_set['Y_current'] = 0.5 * (target_set['Y_org_current'] + target_set['Y_context'])
target_set['w1_input'] = 0.5 * (target_set['w1_org'] + target_set['w1_context'])
else:
raise ValueError('mode not supported: {}'.format(mode))
return target_set
|