File size: 15,043 Bytes
51a7969
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8224d2b
9004b7f
 
 
 
51a7969
9004b7f
 
 
 
e0b30b2
9004b7f
 
51a7969
9004b7f
 
 
51a7969
 
 
cbec9d4
9004b7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
51a7969
9004b7f
 
 
 
 
c9c0d41
9004b7f
 
5096125
9004b7f
 
c9c0d41
9004b7f
 
5096125
9004b7f
 
c9c0d41
9004b7f
 
5096125
9004b7f
 
c9c0d41
9004b7f
 
5096125
9004b7f
 
c9c0d41
9004b7f
 
5096125
9004b7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
51a7969
 
9004b7f
4bc8ae4
 
 
 
9004b7f
51a7969
 
9004b7f
 
51a7969
9004b7f
 
51a7969
9004b7f
 
 
 
51a7969
9004b7f
 
51a7969
9004b7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0b75af5
9004b7f
 
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 json
# import gradio as gr
# import spaces
# import torch
# from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
# from sentence_splitter import SentenceSplitter
# from itertools import product

# # Get the Hugging Face token from environment variable
# hf_token = os.getenv('HF_TOKEN')

# cuda_available = torch.cuda.is_available()
# device = torch.device("cpu" if cuda_available else "cpu")
# print(f"Using device: {device}")

# # Initialize paraphraser model and tokenizer
# paraphraser_model_name = "NoaiGPT/777"
# paraphraser_tokenizer = AutoTokenizer.from_pretrained(paraphraser_model_name, use_auth_token=hf_token)
# paraphraser_model = AutoModelForSeq2SeqLM.from_pretrained(paraphraser_model_name, use_auth_token=hf_token).to(device)

# # Initialize classifier model and tokenizer
# classifier_model_name = "andreas122001/roberta-mixed-detector"
# classifier_tokenizer = AutoTokenizer.from_pretrained(classifier_model_name)
# classifier_model = AutoModelForSequenceClassification.from_pretrained(classifier_model_name).to(device)

# # Initialize sentence splitter
# splitter = SentenceSplitter(language='en')

# def classify_text(text):
#     inputs = classifier_tokenizer(text, return_tensors="pt", truncation=True, max_length=512).to(device)
#     with torch.no_grad():
#         outputs = classifier_model(**inputs)
#     probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
#     predicted_class = torch.argmax(probabilities, dim=-1).item()
#     main_label = classifier_model.config.id2label[predicted_class]
#     main_score = probabilities[0][predicted_class].item()
#     return main_label, main_score

# # @spaces.GPU
# def generate_paraphrases(text, setting, output_format):
#     sentences = splitter.split(text)
#     all_sentence_paraphrases = []
    
#     if setting == 1:
#         num_return_sequences = 5
#         repetition_penalty = 1.1
#         no_repeat_ngram_size = 2
#         temperature = 1.0
#         max_length = 128
#     elif setting == 2:
#         num_return_sequences = 10
#         repetition_penalty = 1.2
#         no_repeat_ngram_size = 3
#         temperature = 1.2
#         max_length = 192
#     elif setting == 3:
#         num_return_sequences = 15
#         repetition_penalty = 1.3
#         no_repeat_ngram_size = 4
#         temperature = 1.4
#         max_length = 256
#     elif setting == 4:
#         num_return_sequences = 20
#         repetition_penalty = 1.4
#         no_repeat_ngram_size = 5
#         temperature = 1.6
#         max_length = 320
#     else:
#         num_return_sequences = 25
#         repetition_penalty = 1.5
#         no_repeat_ngram_size = 6
#         temperature = 1.8
#         max_length = 384
    
#     top_k = 50
#     top_p = 0.95
#     length_penalty = 1.0
    
#     formatted_output = "Original text:\n" + text + "\n\n"
#     formatted_output += "Paraphrased versions:\n"
    
#     json_output = {
#         "original_text": text,
#         "paraphrased_versions": [],
#         "combined_versions": [],
#         "human_like_versions": []
#     }
    
#     for i, sentence in enumerate(sentences):
#         inputs = paraphraser_tokenizer(f'paraphraser: {sentence}', return_tensors="pt", padding="longest", truncation=True, max_length=max_length).to(device)
        
#         # Generate paraphrases using the specified parameters
#         outputs = paraphraser_model.generate(
#             inputs.input_ids,
#             attention_mask=inputs.attention_mask,
#             num_return_sequences=num_return_sequences,
#             repetition_penalty=repetition_penalty,
#             no_repeat_ngram_size=no_repeat_ngram_size,
#             temperature=temperature,
#             max_length=max_length,
#             top_k=top_k,
#             top_p=top_p,
#             do_sample=True,
#             early_stopping=False,
#             length_penalty=length_penalty
#         )
        
#         paraphrases = paraphraser_tokenizer.batch_decode(outputs, skip_special_tokens=True)
        
#         formatted_output += f"Original sentence {i+1}: {sentence}\n"
#         for j, paraphrase in enumerate(paraphrases, 1):
#             formatted_output += f"  Paraphrase {j}: {paraphrase}\n"
        
#         json_output["paraphrased_versions"].append({
#             f"original_sentence_{i+1}": sentence,
#             "paraphrases": paraphrases
#         })
        
#         all_sentence_paraphrases.append(paraphrases)
#         formatted_output += "\n"
    
#     all_combinations = list(product(*all_sentence_paraphrases))
    
#     formatted_output += "\nCombined paraphrased versions:\n"
#     combined_versions = []
#     for i, combination in enumerate(all_combinations[:50], 1):  # Limit to 50 combinations
#         combined_paraphrase = " ".join(combination)
#         combined_versions.append(combined_paraphrase)
    
#     json_output["combined_versions"] = combined_versions
    
#     # Classify combined versions
#     human_versions = []
#     for i, version in enumerate(combined_versions, 1):
#         label, score = classify_text(version)
#         formatted_output += f"Version {i}:\n{version}\n"
#         formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
#         if label == "human-produced" or (label == "machine-generated" and score < 0.98):
#             human_versions.append((version, label, score))
    
#     formatted_output += "\nHuman-like or Less Confident Machine-generated versions:\n"
#     for i, (version, label, score) in enumerate(human_versions, 1):
#         formatted_output += f"Version {i}:\n{version}\n"
#         formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
    
#     json_output["human_like_versions"] = [
#         {"version": version, "label": label, "confidence_score": score}
#         for version, label, score in human_versions
#     ]
    
#     # If no human-like versions, include the top 5 least confident machine-generated versions
#     if not human_versions:
#         human_versions = sorted([(v, l, s) for v, l, s in zip(combined_versions, [classify_text(v)[0] for v in combined_versions], [classify_text(v)[1] for v in combined_versions])], key=lambda x: x[2])[:5]
#         formatted_output += "\nNo human-like versions found. Showing top 5 least confident machine-generated versions:\n"
#         for i, (version, label, score) in enumerate(human_versions, 1):
#             formatted_output += f"Version {i}:\n{version}\n"
#             formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
    
#     if output_format == "text":
#         return formatted_output, "\n\n".join([v[0] for v in human_versions])
#     else:
#         return json.dumps(json_output, indent=2), "\n\n".join([v[0] for v in human_versions])

# # Define the Gradio interface
# iface = gr.Interface(
#     fn=generate_paraphrases,
#     inputs=[
#         gr.Textbox(lines=5, label="Input Text"),
#         gr.Slider(minimum=1, maximum=5, step=1, label="Readability to Human-like Setting"),
#         gr.Radio(["text", "json"], label="Output Format")
#     ],
#     outputs=[
#         gr.Textbox(lines=20, label="Detailed Paraphrases and Classifications"),
#         gr.Textbox(lines=10, label="Human-like or Less Confident Machine-generated Paraphrases")
#     ],
#     title="Advanced Diverse Paraphraser with Human-like Filter",
#     description="Enter a text, select a setting from readable to human-like, and choose the output format to generate diverse paraphrased versions. Combined versions are classified, and those detected as human-produced or less confidently machine-generated are presented in the final output."
# )

# # Launch the interface
# iface.launch()

import os
import json
import gradio as gr
import spaces
import torch
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
from sentence_splitter import SentenceSplitter
from itertools import product

# Get the Hugging Face token from environment variable
hf_token = os.getenv('HF_TOKEN')

cuda_available = torch.cuda.is_available()
device = torch.device("cuda" if cuda_available else "cpu")
print(f"Using device: {device}")

# Initialize paraphraser model and tokenizer
paraphraser_model_name = "sharad/ParaphraseGPT"
paraphraser_tokenizer = AutoTokenizer.from_pretrained("humarin/chatgpt_paraphraser_on_T5_base")
paraphraser_model = AutoModelForSeq2SeqLM.from_pretrained(paraphraser_model_name).to(device)
paraphrase_pipeline = pipeline("text2text-generation", model=paraphraser_model, tokenizer=paraphraser_tokenizer, device=0 if cuda_available else -1)

# Initialize classifier model and tokenizer
classifier_model_name = "andreas122001/roberta-mixed-detector"
classifier_tokenizer = AutoTokenizer.from_pretrained(classifier_model_name)
classifier_model = AutoModelForSequenceClassification.from_pretrained(classifier_model_name).to(device)

# Initialize sentence splitter
splitter = SentenceSplitter(language='en')

def classify_text(text):
    inputs = classifier_tokenizer(text, return_tensors="pt", truncation=True, max_length=512).to(device)
    with torch.no_grad():
        outputs = classifier_model(**inputs)
    probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
    predicted_class = torch.argmax(probabilities, dim=-1).item()
    main_label = classifier_model.config.id2label[predicted_class]
    main_score = probabilities[0][predicted_class].item()
    return main_label, main_score

@spaces.GPU
def generate_paraphrases(text, setting, output_format):
    sentences = splitter.split(text)
    all_sentence_paraphrases = []
    
    if setting == 1:
        num_return_sequences = 5
        repetition_penalty = 1.1
        no_repeat_ngram_size = 2
        temperature = 0.9
        max_length = 128
    elif setting == 2:
        num_return_sequences = 5
        repetition_penalty = 1.2
        no_repeat_ngram_size = 3
        temperature = 0.95
        max_length = 192
    elif setting == 3:
        num_return_sequences = 5
        repetition_penalty = 1.3
        no_repeat_ngram_size = 4
        temperature = 1.0
        max_length = 256
    elif setting == 4:
        num_return_sequences = 5
        repetition_penalty = 1.4
        no_repeat_ngram_size = 5
        temperature = 1.05
        max_length = 320
    else:
        num_return_sequences = 5
        repetition_penalty = 1.5
        no_repeat_ngram_size = 6
        temperature = 1.1
        max_length = 384
    
    top_k = 50
    top_p = 0.95
    length_penalty = 1.0
    
    formatted_output = "Original text:\n" + text + "\n\n"
    formatted_output += "Paraphrased versions:\n"
    
    json_output = {
        "original_text": text,
        "paraphrased_versions": [],
        "combined_versions": [],
        "human_like_versions": []
    }
    
    for i, sentence in enumerate(sentences):
        paraphrases = paraphrase_pipeline(
            sentence,
            num_return_sequences=num_return_sequences,
            do_sample=True,
            top_k=top_k,
            top_p=top_p,
            temperature=temperature,
            no_repeat_ngram_size=no_repeat_ngram_size,
            repetition_penalty=repetition_penalty,
            max_length=max_length
        )
        
        paraphrases_texts = [p['generated_text'] for p in paraphrases]
        
        formatted_output += f"Original sentence {i+1}: {sentence}\n"
        for j, paraphrase in enumerate(paraphrases_texts, 1):
            formatted_output += f"  Paraphrase {j}: {paraphrase}\n"
        
        json_output["paraphrased_versions"].append({
            f"original_sentence_{i+1}": sentence,
            "paraphrases": paraphrases_texts
        })
        
        all_sentence_paraphrases.append(paraphrases_texts)
        formatted_output += "\n"
    
    all_combinations = list(product(*all_sentence_paraphrases))
    
    formatted_output += "\nCombined paraphrased versions:\n"
    combined_versions = []
    for i, combination in enumerate(all_combinations[:50], 1):  # Limit to 50 combinations
        combined_paraphrase = " ".join(combination)
        combined_versions.append(combined_paraphrase)
    
    json_output["combined_versions"] = combined_versions
    
    # Classify combined versions
    human_versions = []
    for i, version in enumerate(combined_versions, 1):
        label, score = classify_text(version)
        formatted_output += f"Version {i}:\n{version}\n"
        formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
        if label == "human-produced" or (label == "machine-generated" and score < 0.98):
            human_versions.append((version, label, score))
    
    formatted_output += "\nHuman-like or Less Confident Machine-generated versions:\n"
    for i, (version, label, score) in enumerate(human_versions, 1):
        formatted_output += f"Version {i}:\n{version}\n"
        formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
    
    json_output["human_like_versions"] = [
        {"version": version, "label": label, "confidence_score": score}
        for version, label, score in human_versions
    ]
    
    # If no human-like versions, include the top 5 least confident machine-generated versions
    if not human_versions:
        human_versions = sorted([(v, l, s) for v, l, s in zip(combined_versions, [classify_text(v)[0] for v in combined_versions], [classify_text(v)[1] for v in combined_versions])], key=lambda x: x[2])[:5]
        formatted_output += "\nNo human-like versions found. Showing top 5 least confident machine-generated versions:\n"
        for i, (version, label, score) in enumerate(human_versions, 1):
            formatted_output += f"Version {i}:\n{version}\n"
            formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
    
    if output_format == "text":
        return formatted_output, "\n\n".join([v[0] for v in human_versions])
    else:
        return json.dumps(json_output, indent=2), "\n\n".join([v[0] for v in human_versions])

# Define the Gradio interface
iface = gr.Interface(
    fn=generate_paraphrases,
    inputs=[
        gr.Textbox(lines=5, label="Input Text"),
        gr.Slider(minimum=1, maximum=5, step=1, label="Readability to Human-like Setting"),
        gr.Radio(["text", "json"], label="Output Format")
    ],
    outputs=[
        gr.Textbox(lines=20, label="Detailed Paraphrases and Classifications"),
        gr.Textbox(lines=10, label="Human-like or Less Confident Machine-generated Paraphrases")
    ],
    title="Advanced Diverse Paraphraser with Human-like Filter",
    description="Enter a text, select a setting from readable to human-like, and choose the output format to generate diverse paraphrased versions. Combined versions are classified, and those detected as human-produced or less confidently machine-generated are presented in the final output."
)

# Launch the interface
iface.launch()