File size: 12,059 Bytes
45d10c4
 
 
 
 
 
 
 
 
a00beed
d22f052
79b97e2
 
 
 
 
 
 
 
 
45d10c4
 
 
 
 
d22f052
edc836f
 
45d10c4
 
 
 
 
 
 
79b97e2
 
45d10c4
 
 
 
 
 
 
 
 
 
 
 
79b97e2
e81407b
 
9df8406
dd9b08a
a00beed
 
 
d22f052
dd9b08a
350b1a0
 
 
a00beed
 
d22f052
 
 
 
 
 
 
a00beed
75ba191
 
 
 
 
 
 
 
 
 
b472976
a00beed
 
45d10c4
 
79b97e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f45e494
 
 
 
 
 
 
 
79b97e2
 
 
 
 
dd9b08a
 
 
 
 
 
 
 
 
 
 
45d10c4
 
 
dd9b08a
45d10c4
 
dd9b08a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
45d10c4
dd9b08a
 
 
 
 
 
 
 
 
 
 
 
45d10c4
dd9b08a
 
45d10c4
dd9b08a
 
 
45d10c4
 
f45e494
 
 
45d10c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9df8406
 
 
 
a00beed
 
9df8406
 
dd9b08a
9df8406
 
 
 
6f614b5
9df8406
 
 
 
 
 
 
 
d22f052
9df8406
6f614b5
 
 
 
 
45d10c4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
350b1a0
 
 
45d10c4
 
ff03afa
45d10c4
 
 
 
 
350b1a0
 
 
1c49ee1
 
 
 
 
79b97e2
45d10c4
 
 
9fc992f
8fe6e3e
 
350b1a0
 
 
8fe6e3e
 
 
 
 
 
 
 
350b1a0
 
 
1c49ee1
 
 
 
 
9fc992f
 
 
 
 
 
 
 
 
 
 
2333c59
9fc992f
 
 
 
 
 
 
8fe6e3e
9fc992f
 
 
 
 
8fe6e3e
 
 
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
import torch
import numpy as np
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import nltk
import torch.nn.functional as F
import nltk
from scipy.special import softmax
import yaml
from utils import *
import joblib
from optimum.bettertransformer import BetterTransformer
import gc
from cleantext import clean
import gradio as gr
from tqdm.auto import tqdm
from transformers import pipeline
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import nltk
from nltk.tokenize import sent_tokenize
from optimum.pipelines import pipeline

with open("config.yaml", "r") as file:
    params = yaml.safe_load(file)
nltk.download("punkt")
nltk.download("stopwords")
device_needed = "cuda" if torch.cuda.is_available() else "cpu"
device = "cuda" if torch.cuda.is_available() else "cpu"

text_bc_model_path = params["TEXT_BC_MODEL_PATH"]
text_mc_model_path = params["TEXT_MC_MODEL_PATH"]
text_quillbot_model_path = params["TEXT_QUILLBOT_MODEL_PATH"]
quillbot_labels = params["QUILLBOT_LABELS"]
mc_label_map = params["MC_OUTPUT_LABELS"]
mc_token_size = int(params["MC_TOKEN_SIZE"])
bc_token_size = int(params["BC_TOKEN_SIZE"])
bias_checker_model_name = params['BIAS_CHECKER_MODEL_PATH']
bias_corrector_model_name = params['BIAS_CORRECTOR_MODEL_PATH']
text_bc_tokenizer = AutoTokenizer.from_pretrained(text_bc_model_path)
text_bc_model = AutoModelForSequenceClassification.from_pretrained(
    text_bc_model_path
).to(device)
text_mc_tokenizer = AutoTokenizer.from_pretrained(text_mc_model_path)
text_mc_model = AutoModelForSequenceClassification.from_pretrained(
    text_mc_model_path
).to(device)
quillbot_tokenizer = AutoTokenizer.from_pretrained(text_quillbot_model_path)
quillbot_model = AutoModelForSequenceClassification.from_pretrained(
    text_quillbot_model_path
).to(device)



# proxy models for explainability
mini_bc_model_name = "polygraf-ai/bc-model"
bc_tokenizer_mini = AutoTokenizer.from_pretrained(mini_bc_model_name)
bc_model_mini = AutoModelForSequenceClassification.from_pretrained(
    mini_bc_model_name
).to(device_needed)
mini_humanizer_model_name =  "polygraf-ai/humanizer-model"
humanizer_tokenizer_mini = AutoTokenizer.from_pretrained(
    mini_humanizer_model_name
)
humanizer_model_mini = AutoModelForSequenceClassification.from_pretrained(
    mini_humanizer_model_name
).to(device_needed)

bc_model_mini = BetterTransformer.transform(bc_model_mini)
humanizer_model_mini = BetterTransformer.transform(humanizer_model_mini)
text_bc_model = BetterTransformer.transform(text_bc_model)
text_mc_model = BetterTransformer.transform(text_mc_model)
quillbot_model = BetterTransformer.transform(quillbot_model)

bias_model_checker = AutoModelForSequenceClassification.from_pretrained(bias_checker_model_name)
tokenizer = AutoTokenizer.from_pretrained(bias_checker_model_name)
bias_model_checker = BetterTransformer.transform(bias_model_checker, keep_original_model=False)
bias_checker = pipeline(
    "text-classification",
    model=bias_checker_model_name,
    tokenizer=bias_checker_model_name,
)
gc.collect()
bias_corrector = pipeline( "text2text-generation", model=bias_corrector_model_name, accelerator="ort")

# model score calibration
iso_reg = joblib.load("isotonic_regression_model.joblib")


def split_text(text: str) -> list:
    sentences = sent_tokenize(text)
    return [[sentence] for sentence in sentences]

def correct_text(text: str, bias_checker, bias_corrector, separator: str = " ") -> tuple:
    sentence_batches = split_text(text)
    corrected_text = []
    corrections = []
    for batch in tqdm(sentence_batches, total=len(sentence_batches), desc="correcting text.."):
        raw_text = " ".join(batch)
        results = bias_checker(raw_text)
        if results[0]["label"] != "LABEL_1" or (results[0]["label"] == "LABEL_1" and results[0]["score"] < 0.9):
            corrected_batch = bias_corrector(raw_text)
            corrected_version = corrected_batch[0]["generated_text"]
            corrected_text.append(corrected_version)
            corrections.append((raw_text, corrected_version)) 
        else:
            corrected_text.append(raw_text)
    corrected_text = separator.join(corrected_text)
    return corrected_text, corrections

def update(text: str):
    text = clean(text, lower=False)
    corrected_text, corrections = correct_text(text, bias_checker, bias_corrector)
    corrections_display = "".join([f"{corr}" for orig, corr in corrections])
    if corrections_display == "":
        corrections_display = text
    return corrections_display

def update_main(text: str):
    text = clean(text, lower=False)
    corrected_text, corrections = correct_text(text, bias_checker, bias_corrector)
    corrections_display = "\n\n".join([f"Original: {orig}\nCorrected: {corr}" for orig, corr in corrections])
    return corrected_text, corrections_display

def split_text(text: str) -> list:
    sentences = sent_tokenize(text)
    return [[sentence] for sentence in sentences]

def get_token_length(tokenizer, sentence):
    return len(tokenizer.tokenize(sentence))

def split_text_allow_complete_sentences_nltk(text, type_det="bc"):
    sentences = sent_tokenize(text)
    chunks = []
    current_chunk = []
    current_length = 0
    if type_det == "bc":
        tokenizer = text_bc_tokenizer
        max_tokens = bc_token_size
    elif type_det == "mc":
        tokenizer = text_mc_tokenizer
        max_tokens = mc_token_size

    elif type_det == "quillbot":
        tokenizer = quillbot_tokenizer
        max_tokens = 256

    def add_sentence_to_chunk(sentence):
        nonlocal current_chunk, current_length
        sentence_length = get_token_length(tokenizer, sentence)
        if current_length + sentence_length > max_tokens:
            chunks.append((current_chunk, current_length))
            current_chunk = []
            current_length = 0
        current_chunk.append(sentence)
        current_length += sentence_length

    for sentence in sentences:
        add_sentence_to_chunk(sentence)
    if current_chunk:
        chunks.append((current_chunk, current_length))
    adjusted_chunks = []
    while chunks:
        chunk = chunks.pop(0)
        if len(chunks) > 0 and chunk[1] < max_tokens / 2:
            next_chunk = chunks.pop(0)
            combined_length = chunk[1] + next_chunk[1]
            if combined_length <= max_tokens:
                adjusted_chunks.append((chunk[0] + next_chunk[0], combined_length))
            else:
                adjusted_chunks.append(chunk)
                chunks.insert(0, next_chunk)
        else:
            adjusted_chunks.append(chunk)
    result_chunks = [" ".join(chunk[0]) for chunk in adjusted_chunks]
    return result_chunks


def predict_quillbot(text, bias_buster_selected):
    if bias_buster_selected:
        text = update(text)
    with torch.no_grad():
        quillbot_model.eval()
        tokenized_text = quillbot_tokenizer(
            text,
            padding="max_length",
            truncation=True,
            max_length=256,
            return_tensors="pt",
        ).to(device)
        output = quillbot_model(**tokenized_text)
        output_norm = softmax(output.logits.detach().cpu().numpy(), 1)[0]
        q_score = {
            "Humanized": output_norm[1].item(),
            "Original": output_norm[0].item(),
        }
        return q_score


def predict_for_explainanility(text, model_type=None):
    if model_type == "quillbot":
        cleaning = False
        max_length = 256
        model = humanizer_model_mini
        tokenizer = humanizer_tokenizer_mini
    elif model_type == "bc":
        cleaning = True
        max_length = bc_token_size
        model = bc_model_mini
        tokenizer = bc_tokenizer_mini
    else:
        raise ValueError("Invalid model type")
    with torch.no_grad():
        if cleaning:
            text = [remove_special_characters(t) for t in text]
        tokenized_text = tokenizer(
            text,
            return_tensors="pt",
            padding="max_length",
            truncation=True,
            max_length=max_length,
        ).to(device_needed)
        outputs = model(**tokenized_text)
        tensor_logits = outputs[0]
        probas = F.softmax(tensor_logits).detach().cpu().numpy()
    return probas


def predict_bc(model, tokenizer, text):
    with torch.no_grad():
        model.eval()
        tokens = text_bc_tokenizer(
            text,
            padding="max_length",
            truncation=True,
            max_length=bc_token_size,
            return_tensors="pt",
        ).to(device)
        output = model(**tokens)
        output_norm = softmax(output.logits.detach().cpu().numpy(), 1)[0]
        return output_norm


def predict_mc(model, tokenizer, text):
    with torch.no_grad():
        model.eval()
        tokens = text_mc_tokenizer(
            text,
            padding="max_length",
            truncation=True,
            return_tensors="pt",
            max_length=mc_token_size,
        ).to(device)
        output = model(**tokens)
        output_norm = softmax(output.logits.detach().cpu().numpy(), 1)[0]
        return output_norm


def predict_bc_scores(input):
    bc_scores = []
    samples_len_bc = len(
        split_text_allow_complete_sentences_nltk(input, type_det="bc")
    )
    segments_bc = split_text_allow_complete_sentences_nltk(input, type_det="bc")
    for i in range(samples_len_bc):
        cleaned_text_bc = remove_special_characters(segments_bc[i])
        bc_score = predict_bc(text_bc_model, text_bc_tokenizer, cleaned_text_bc)
        bc_scores.append(bc_score)
    bc_scores_array = np.array(bc_scores)
    average_bc_scores = np.mean(bc_scores_array, axis=0)
    bc_score_list = average_bc_scores.tolist()
    print(
        f"Original BC scores: AI: {bc_score_list[1]}, HUMAN: {bc_score_list[0]}"
    )
    # isotonic regression calibration
    ai_score = iso_reg.predict([bc_score_list[1]])[0]
    human_score = 1 - ai_score
    bc_score = {"AI": ai_score, "HUMAN": human_score}
    print(f"Calibration BC scores: AI: {ai_score}, HUMAN: {human_score}")
    print(f"Input Text: {cleaned_text_bc}")
    return bc_score


def predict_mc_scores(input):
    # BC SCORE
    bc_scores = []
    samples_len_bc = len(
        split_text_allow_complete_sentences_nltk(input, type_det="bc")
    )
    segments_bc = split_text_allow_complete_sentences_nltk(input, type_det="bc")
    for i in range(samples_len_bc):
        cleaned_text_bc = remove_special_characters(segments_bc[i])
        bc_score = predict_bc(text_bc_model, text_bc_tokenizer, cleaned_text_bc)
        bc_scores.append(bc_score)
    bc_scores_array = np.array(bc_scores)
    average_bc_scores = np.mean(bc_scores_array, axis=0)
    bc_score_list = average_bc_scores.tolist()
    print(
        f"Original BC scores: AI: {bc_score_list[1]}, HUMAN: {bc_score_list[0]}"
    )
    # isotonic regression calibration
    ai_score = iso_reg.predict([bc_score_list[1]])[0]
    human_score = 1 - ai_score
    bc_score = {"AI": ai_score, "HUMAN": human_score}
    print(f"Calibration BC scores: AI: {ai_score}, HUMAN: {human_score}")
    mc_scores = []
    segments_mc = split_text_allow_complete_sentences_nltk(
        input, type_det="mc"
    )
    samples_len_mc = len(
        split_text_allow_complete_sentences_nltk(input, type_det="mc")
    )
    for i in range(samples_len_mc):
        cleaned_text_mc = remove_special_characters(segments_mc[i])
        mc_score = predict_mc(
            text_mc_model, text_mc_tokenizer, cleaned_text_mc
        )
        mc_scores.append(mc_score)
    mc_scores_array = np.array(mc_scores)
    average_mc_scores = np.mean(mc_scores_array, axis=0)
    mc_score_list = average_mc_scores.tolist()
    mc_score = {}
    for score, label in zip(mc_score_list, mc_label_map):
        mc_score[label.upper()] = score

    sum_prob = 1 - bc_score["HUMAN"]
    for key, value in mc_score.items():
        mc_score[key] = value * sum_prob
    print("MC Score:", mc_score)
    if sum_prob < 0.01:
        mc_score = {}

    return mc_score