-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_marabou.py
331 lines (273 loc) · 12.6 KB
/
run_marabou.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
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
import argparse
import multiprocessing
from pathlib import Path
import evaluate
import numpy as np
import pandas as pd
import yaml
from datasets import load_dataset
from nltk.tokenize import sent_tokenize, word_tokenize
from tqdm import tqdm
import nltk
nltk.download('punkt')
import settings
def load_categories(file_path):
"""Load categories dictionary from a YAML file."""
with open(file_path, 'r', encoding='utf-8') as yaml_file:
return yaml.safe_load(yaml_file)
def worker(worker_id, data_slice, categories_dict, sentence_threshold,
prompt_length, results_queue):
results = []
for text in tqdm(data_slice, desc=f"Process {worker_id}"):
results.extend(
process_text(text, categories_dict, sentence_threshold,
prompt_length))
results_queue.put(results)
def worker_prompt(worker_id, prompts, sentences, categories_dict, toxicitys,
results_queue):
results = []
for prompt, sentence, toxicity in tqdm(zip(prompts, sentences, toxicitys),
total=len(prompts),
desc=f"Process {worker_id}"):
original_words = word_tokenize(prompt)
prompt = ' '.join(original_words)
results.extend(
classify_prompt_based_on_category(prompt, sentence,
categories_dict, toxicity))
results_queue.put(results)
def process_text(text, categories_dict, sentence_threshold, prompt_length):
"""Process text to extract prompts based on categories."""
results = []
sentences = sent_tokenize(text) # Split text into sentences
for sentence in sentences:
original_words = word_tokenize(sentence) # Tokenize the sentence
if len(original_words) >= sentence_threshold:
prompt = ' '.join(
original_words[:min(prompt_length, len(original_words))])
results.extend(
classify_prompt_based_on_category(prompt, sentence,
categories_dict))
return results
def classify_prompt_based_on_category(prompt: str, sentence: str,
categories_dict: dict) -> list:
"""
Classify the given prompt into categories based on the provided categories dictionary.
:param prompt: The prompt string to classify.
:param categories_dict: A dictionary with domain as keys and another dictionary as values,
where the inner dictionary has category as keys and list of keywords as values.
:return: A list of dictionaries, where each dictionary contains the domain, category, and the prompt.
"""
results = [] # List to store the result
# Convert prompt to lowercase and split into words
words_lower = prompt.lower().split()
# Generate possible phrases from the prompt words
possible_phrases_lower = [
' '.join(words_lower[i:j + 1]) for i in range(len(words_lower))
for j in range(i, min(i + 2, len(words_lower)))
]
# Iterate through categories dictionary to find matching categories
for domain, categories in categories_dict.items():
flag = 0
for category, keywords in categories.items():
for keyword in keywords:
keyword_lower = keyword.lower()
# Check if the keyword matches any phrase from the prompt
if any(keyword_lower == phrase
for phrase in possible_phrases_lower):
flag += 1
new_row = {
"domain": domain,
"category": category,
"texts": sentence,
"prompts": prompt,
}
break
if flag > 1:
break
if flag == 1:
results.append(new_row)
return results
def load_dataset_chunks(dataset_name, num_processes):
"""Load the specified dataset and split it into chunks for multiprocessing."""
if dataset_name == "cnn_dailymail":
dataset = load_dataset("ccdv/cnn_dailymail", '3.0.0')['train']
data = dataset.to_pandas()
data_chunks = np.array_split(data['article'], num_processes)
file_name = "cnn_dailymail_new.jsonl"
elif dataset_name == "bookcorpus":
dataset = load_dataset('bookcorpus')['train']
data = dataset.to_pandas()
data_chunks = np.array_split(data['text'], num_processes)
file_name = "bookcorpus_new.jsonl"
else:
raise ValueError("Unknown dataset")
return data_chunks, file_name
def post_process(data: pd.DataFrame):
max_samples = 3000
random_state = 42
all_data = pd.DataFrame()
unique_domains = data['domain'].unique()
for domain in unique_domains:
domain_data = data[data['domain'] == domain]
unique_categories = domain_data['category'].unique()
domain_balanced = []
for category in unique_categories:
category_data = domain_data[domain_data['category'] == category]
sample_size = min(len(category_data),
int(max_samples / len(unique_categories)))
if len(category_data) > 0:
sampled_data = category_data.sample(n=sample_size,
random_state=random_state)
domain_balanced.append(sampled_data)
all_data = pd.concat([all_data, pd.concat(domain_balanced)])
return all_data
def calculate_scores(df, toxicity_evaluator, regard_evaluator):
input_texts = df['prompts'].tolist()
toxicity_results = toxicity_evaluator.compute(predictions=input_texts)
toxicity_scores = toxicity_results["toxicity"]
regard_results = regard_evaluator.compute(data=input_texts)
regard_scores = [
d[0]['score'] for d in regard_results['regard'] for l in d
if l['label'] == 'negative'
]
df['toxicity'] = toxicity_scores
df['regard'] = regard_scores
df['total_score'] = df['toxicity'] + df['regard']
return df
def select_top_1k_per_category(df, file_path):
toxicity_evaluator = evaluate.load("toxicity", module_type="measurement")
regard_evaluator = evaluate.load("regard", module_type="measurement")
df = calculate_scores(df, toxicity_evaluator, regard_evaluator)
domain_category_counts = df.groupby(
'domain')['category'].nunique().reset_index(name='unique_categories')
domain_category_counts['per_category_quota'] = 1000
df = df.merge(domain_category_counts[['domain', 'per_category_quota']],
on='domain')
print(f"Intermediate results are stored in {file_path}")
df.to_json(file_path, orient='records', lines=True)
def select_by_quota(group_df):
quota = int(group_df['per_category_quota'].iloc[0])
return group_df.nlargest(min(quota, len(group_df)), 'total_score')
top_2k_per_domain = df.groupby(['domain', 'category'],
group_keys=False).apply(select_by_quota)
return top_2k_per_domain
def resample_large_domains(dataframe, threshold=60000, random_seed=42):
"""
"""
dataframe = dataframe.drop_duplicates(subset='prompts').reset_index(drop=True)
domain_counts = dataframe['domain'].value_counts()
large_domains = domain_counts[domain_counts > threshold].index
filtered_df = dataframe[dataframe['domain'].isin(large_domains)]
def resample_domain_group(group, target_size, seed):
category_counts = group['category'].value_counts(normalize=True)
samples_per_category = (target_size * category_counts).astype(int)
resampled_group = group.groupby('category', group_keys=False).apply(
lambda x: x.sample(
n=min(len(x), samples_per_category.loc[x.name]),
random_state=seed,
replace=False # Use replace=False to avoid duplicates
)
)
return resampled_group
resampled_df = filtered_df.groupby('domain', group_keys=False).apply(
lambda x: resample_domain_group(x, threshold, random_seed)
).reset_index(drop=True)
cleaned_df = dataframe[~dataframe['domain'].isin(large_domains)]
final_df = pd.concat([cleaned_df, resampled_df], ignore_index=True)
return final_df
def main():
# TODO: fix domain and category
parser = argparse.ArgumentParser(
description='Process text to extract prompts based on categories.')
parser.add_argument('--sentence_threshold',
type=int,
default=20,
help='Sentence length threshold')
parser.add_argument('--prompt_length',
type=int,
default=10,
help='Prompt length in terms of number of words')
parser.add_argument('--num_processes',
type=int,
default=60,
help='Number of processes to use')
parser.add_argument('--dataset',
type=str,
default='cnn_dailymail',
help='Name of the dataset to use')
parser.add_argument('--mode',
type=str,
default='text',
choices=['text', 'prompt'],
help='Mode of classification: text or prompt')
args = parser.parse_args()
wordlists_dir = settings.WORDLIST_DIR
gender_dict = load_categories(wordlists_dir / "gender.yaml")
religion_dict = load_categories(wordlists_dir / "religion.yaml")
age_dict = load_categories(wordlists_dir / "age.yaml")
race_dict = load_categories(wordlists_dir / "race.yaml")
bodyshaming_dict = load_categories(wordlists_dir / "bodyshaming.yaml")
socioeconomic_dict = load_categories(wordlists_dir / "socioeconomic.yaml")
lgbt_dict = load_categories(wordlists_dir / "lgbt.yaml")
appearance_dict = load_categories(wordlists_dir / "appearance.yaml")
class_dict = load_categories(wordlists_dir / "class.yaml")
education_dict = load_categories(wordlists_dir / "education.yaml")
disability_dict = load_categories(wordlists_dir / "disability.yaml")
national_dict = load_categories(wordlists_dir / "national.yaml")
categories_dict = {
'gender': gender_dict,
'religion': religion_dict,
'age': age_dict,
'race': race_dict,
'bodyshaming': bodyshaming_dict,
'socioeconomic': socioeconomic_dict,
'lgbt': lgbt_dict,
'appearance': appearance_dict,
'class': class_dict,
'education': education_dict,
'disability': disability_dict,
'national': national_dict
}
data_chunks, file_name = load_dataset_chunks(args.dataset,
args.num_processes)
file_path = Path(settings.DATASET_PATH) / file_name
# file_path.mkdir(parents=True, exist_ok=True)
# if file_path.exists():
# print(f"File {file_path}'s parent dir has been created successfully.")
# else:
# print(f"Failed to create the parent dir of file {file_path}.")
manager = multiprocessing.Manager()
results_queue = manager.Queue()
processes = []
for i, data_chunk in enumerate(data_chunks):
if args.mode == 'text':
p = multiprocessing.Process(
target=worker,
args=(i, data_chunk, categories_dict, args.sentence_threshold,
args.prompt_length, results_queue))
elif args.mode == 'prompt':
p = multiprocessing.Process(
target=worker_prompt,
args=(i, data_chunk['prompts'], data_chunk['texts'],
categories_dict, data_chunk['toxicity'], results_queue))
processes.append(p)
p.start()
results = []
for _ in range(args.num_processes):
results.extend(results_queue.get())
for p in processes:
p.join()
results_df = pd.DataFrame(results)
print(results_df.groupby(['domain']).size())
print(results_df.groupby(['domain', 'category']).size())
results_df = resample_large_domains(results_df)
print(results_df.groupby(['domain']).size())
print(results_df.groupby(['domain', 'category']).size())
if args.mode == "text":
results_df = select_top_1k_per_category(results_df, Path(settings.DATASET_PATH) / ("intermediate_result" + file_name))
print(results_df.groupby(['domain']).size())
print(results_df.groupby(['domain', 'category']).size())
results_df.to_json(file_path, orient='records', lines=True)
print(f"Results saved to {file_path}")
if __name__ == "__main__":
main()