File size: 19,225 Bytes
6ca0d72
 
 
7eaaff0
ad52429
6ca0d72
ad52429
6ca0d72
25f2b88
7eaaff0
 
6ca0d72
256b1d4
25f2b88
 
6fc3054
 
41365d5
6ca0d72
7eaaff0
 
 
6ca0d72
 
 
 
 
 
 
 
 
 
 
ad52429
7eaaff0
6ca0d72
7eaaff0
 
 
6ca0d72
7eaaff0
ad52429
7eaaff0
 
6ca0d72
7eaaff0
 
 
6ca0d72
7eaaff0
 
25f2b88
6ca0d72
25f2b88
7eaaff0
6ca0d72
7eaaff0
 
 
 
6ca0d72
7b9d8b2
 
6ca0d72
7b9d8b2
 
 
 
 
 
 
 
6ca0d72
 
25f2b88
7eaaff0
 
1bb7d9d
7eaaff0
 
 
 
1bb7d9d
7eaaff0
 
 
1bb7d9d
7eaaff0
 
6ca0d72
 
ad52429
7eaaff0
6ca0d72
5f61427
 
 
 
6ca0d72
53f5f55
ad52429
7eaaff0
 
ad52429
 
25f2b88
6ca0d72
ad52429
 
 
6ca0d72
ad52429
 
 
 
 
6ca0d72
ad52429
 
6ca0d72
3fe982e
ad52429
0fc43d5
3fe982e
25f2b88
6ca0d72
154774e
ad52429
 
3fe982e
 
ad52429
 
 
6ca0d72
ad52429
7b9d8b2
6ca0d72
 
7b9d8b2
ad52429
 
 
 
 
 
7b9d8b2
6ca0d72
7b9d8b2
 
 
 
ad52429
6ca0d72
7b9d8b2
 
7eaaff0
6ca0d72
7b9d8b2
 
 
 
 
 
 
 
6ca0d72
7b9d8b2
 
 
 
6ca0d72
7b9d8b2
 
 
 
 
 
 
 
 
6ca0d72
25f2b88
 
 
 
7b9d8b2
7eaaff0
7b9d8b2
 
 
7eaaff0
 
6ca0d72
7b9d8b2
 
 
 
 
 
 
 
 
6ca0d72
 
444f8bc
99608c9
 
33fd63d
9a1a827
33fd63d
 
 
 
 
 
 
 
ad52429
444f8bc
6ca0d72
33fd63d
 
 
444f8bc
6ca0d72
33fd63d
444f8bc
6ca0d72
33fd63d
 
444f8bc
6ca0d72
444f8bc
 
 
 
6ca0d72
33fd63d
 
 
 
 
 
 
444f8bc
6ca0d72
33fd63d
 
 
444f8bc
6ca0d72
33fd63d
 
6ca0d72
 
444f8bc
6ca0d72
33fd63d
 
 
 
 
444f8bc
6ca0d72
33fd63d
 
 
444f8bc
6ca0d72
33fd63d
 
 
 
 
444f8bc
6ca0d72
33fd63d
5f61427
 
 
 
444f8bc
33fd63d
 
 
444f8bc
6ca0d72
33fd63d
6ca0d72
33fd63d
 
 
 
444f8bc
33fd63d
 
 
 
 
 
 
444f8bc
6ca0d72
33fd63d
 
 
 
 
 
1bb7d9d
6ca0d72
ad52429
 
 
7eaaff0
 
 
ad52429
6ca0d72
7eaaff0
 
6ca0d72
7eaaff0
6ca0d72
7eaaff0
 
6ca0d72
7eaaff0
6ca0d72
ad52429
 
 
 
7eaaff0
6ca0d72
ad52429
7eaaff0
 
 
 
 
 
 
6ca0d72
7eaaff0
 
 
 
 
 
 
 
 
 
 
 
 
6ca0d72
 
41365d5
 
 
 
 
6ca0d72
41365d5
 
 
 
 
6ca0d72
41365d5
 
 
 
 
 
 
 
6ca0d72
41365d5
 
 
 
 
 
6ca0d72
41365d5
 
 
 
6ca0d72
41365d5
 
6ca0d72
41365d5
 
 
 
 
 
 
 
6ca0d72
41365d5
 
 
 
 
 
 
 
 
 
6ca0d72
 
 
41365d5
6fc3054
6ca0d72
1419b33
ad52429
6ca0d72
5f61427
 
 
 
 
 
 
 
 
 
 
 
 
050caf8
5f61427
 
6ca0d72
 
 
5f61427
 
 
 
 
 
 
6ca0d72
5f61427
 
6ca0d72
5f61427
 
e3d20ad
5f61427
 
 
 
 
6ca0d72
7eaaff0
6ca0d72
d946b22
6ca0d72
 
 
967f5dd
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
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
# AI Text Detector Code Analysis

# IMPORTS AND CONFIGURATION
import torch
import numpy as np
from transformers import AutoTokenizer, AutoModelForSequenceClassification  # HuggingFace transformers for NLP models
import torch.nn.functional as F
import spacy  # Used for sentence splitting
from typing import List, Dict, Tuple
import logging
import os
import gradio as gr  # Used for creating the web UI
from fastapi.middleware.cors import CORSMiddleware
from concurrent.futures import ThreadPoolExecutor
from functools import partial
import time
from datetime import datetime

# Basic logging setup
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# GLOBAL PARAMETERS
MAX_LENGTH = 512  # Maximum token length for the model input
MODEL_NAME = "microsoft/deberta-v3-small"  # Using Microsoft's DeBERTa v3 small model as the base
WINDOW_SIZE = 6  # Number of sentences in each analysis window
WINDOW_OVERLAP = 2  # Number of sentences that overlap between adjacent windows
CONFIDENCE_THRESHOLD = 0.65  # Threshold for highlighting predictions with stronger colors
BATCH_SIZE = 8  # Number of windows to process in a single batch for efficiency
MAX_WORKERS = 4  # Maximum number of worker threads for parallel processing

# TEXT WINDOW PROCESSOR
# This class handles sentence splitting and window creation for text analysis
class TextWindowProcessor:
    def __init__(self):
        # Initialize SpaCy with minimal pipeline for sentence splitting
        try:
            self.nlp = spacy.load("en_core_web_sm")
        except OSError:
            # Auto-download SpaCy model if not available
            logger.info("Downloading spacy model...")
            spacy.cli.download("en_core_web_sm")
            self.nlp = spacy.load("en_core_web_sm")

        # Add sentencizer if not already present
        if 'sentencizer' not in self.nlp.pipe_names:
            self.nlp.add_pipe('sentencizer')

        # Disable unnecessary components for better performance
        disabled_pipes = [pipe for pipe in self.nlp.pipe_names if pipe != 'sentencizer']
        self.nlp.disable_pipes(*disabled_pipes)
        
        # Setup ThreadPoolExecutor for parallel processing
        self.executor = ThreadPoolExecutor(max_workers=MAX_WORKERS)

    # Split text into individual sentences using SpaCy
    def split_into_sentences(self, text: str) -> List[str]:
        doc = self.nlp(text)
        return [str(sent).strip() for sent in doc.sents]

    # Create overlapping windows of fixed size (for quick scan)
    def create_windows(self, sentences: List[str], window_size: int, overlap: int) -> List[str]:
        if len(sentences) < window_size:
            return [" ".join(sentences)]  # Return single window if not enough sentences

        windows = []
        stride = window_size - overlap
        for i in range(0, len(sentences) - window_size + 1, stride):
            window = sentences[i:i + window_size]
            windows.append(" ".join(window))
        return windows

    # Create windows centered around each sentence (for detailed scan)
    # This provides better analysis of individual sentences with proper context
    def create_centered_windows(self, sentences: List[str], window_size: int) -> Tuple[List[str], List[List[int]]]:
        windows = []
        window_sentence_indices = []

        for i in range(len(sentences)):
            half_window = window_size // 2
            start_idx = max(0, i - half_window)
            end_idx = min(len(sentences), i + half_window + 1)

            window = sentences[start_idx:end_idx]
            windows.append(" ".join(window))
            window_sentence_indices.append(list(range(start_idx, end_idx)))

        return windows, window_sentence_indices

# TEXT CLASSIFIER
# This class handles the actual AI/Human classification using a pre-trained model
class TextClassifier:
    def __init__(self):
        # Configure CPU threading if CUDA not available
        if not torch.cuda.is_available():
            torch.set_num_threads(MAX_WORKERS)
            torch.set_num_interop_threads(MAX_WORKERS)
            
        # Set device (GPU if available, otherwise CPU)
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.model_name = MODEL_NAME
        self.tokenizer = None
        self.model = None
        self.processor = TextWindowProcessor()
        self.initialize_model()

    # Initialize the model and tokenizer
    def initialize_model(self):
        logger.info("Initializing model and tokenizer...")
        
        # Using DeBERTa tokenizer specifically for better compatibility
        from transformers import DebertaV2TokenizerFast
        
        self.tokenizer = DebertaV2TokenizerFast.from_pretrained(
            self.model_name,
            model_max_length=MAX_LENGTH,
            use_fast=True  # Use fast tokenizer for better performance
        )
        
        # Load classification model with 2 labels (AI and Human)
        self.model = AutoModelForSequenceClassification.from_pretrained(
            self.model_name,
            num_labels=2
        ).to(self.device)
            
        # Try to load custom fine-tuned model weights if available
        model_path = "model_20250209_184929_acc1.0000.pt"
        if os.path.exists(model_path):
            logger.info(f"Loading custom model from {model_path}")
            checkpoint = torch.load(model_path, map_location=self.device)
            self.model.load_state_dict(checkpoint['model_state_dict'])
        else:
            logger.warning("Custom model file not found. Using base model.")
            
        # Set model to evaluation mode
        self.model.eval()

    # Quick scan analysis - faster but less detailed
    # Uses fixed-size windows with overlap
    def quick_scan(self, text: str) -> Dict:
        if not text.strip():
            return {
                'prediction': 'unknown',
                'confidence': 0.0,
                'num_windows': 0
            }

        # Split text into sentences and then into windows
        sentences = self.processor.split_into_sentences(text)
        windows = self.processor.create_windows(sentences, WINDOW_SIZE, WINDOW_OVERLAP)

        predictions = []
        
        # Process windows in batches for efficiency
        for i in range(0, len(windows), BATCH_SIZE):
            batch_windows = windows[i:i + BATCH_SIZE]

            # Tokenize and prepare input for the model
            inputs = self.tokenizer(
                batch_windows,
                truncation=True,
                padding=True,
                max_length=MAX_LENGTH,
                return_tensors="pt"
            ).to(self.device)

            # Run inference with no gradient calculation
            with torch.no_grad():
                outputs = self.model(**inputs)
                probs = F.softmax(outputs.logits, dim=-1)

                # Process predictions for each window
                for idx, window in enumerate(batch_windows):
                    prediction = {
                        'window': window,
                        'human_prob': probs[idx][1].item(),
                        'ai_prob': probs[idx][0].item(),
                        'prediction': 'human' if probs[idx][1] > probs[idx][0] else 'ai'
                    }
                    predictions.append(prediction)

            # Clean up to free memory
            del inputs, outputs, probs
            if torch.cuda.is_available():
                torch.cuda.empty_cache()

        if not predictions:
            return {
                'prediction': 'unknown',
                'confidence': 0.0,
                'num_windows': 0
            }

        # Average probabilities across all windows for final prediction
        avg_human_prob = sum(p['human_prob'] for p in predictions) / len(predictions)
        avg_ai_prob = sum(p['ai_prob'] for p in predictions) / len(predictions)

        return {
            'prediction': 'human' if avg_human_prob > avg_ai_prob else 'ai',
            'confidence': max(avg_human_prob, avg_ai_prob),
            'num_windows': len(predictions)
        }

    # Detailed scan analysis - slower but provides sentence-level insights
    # Uses windows centered around each sentence for more precise analysis
    def detailed_scan(self, text: str) -> Dict:
        text = text.rstrip()
        
        if not text.strip():
            return {
                'sentence_predictions': [],
                'highlighted_text': '',
                'full_text': '',
                'overall_prediction': {
                    'prediction': 'unknown',
                    'confidence': 0.0,
                    'num_sentences': 0
                }
            }

        # Split text into sentences
        sentences = self.processor.split_into_sentences(text)
        if not sentences:
            return {}

        # Create a window centered on each sentence
        windows, window_sentence_indices = self.processor.create_centered_windows(sentences, WINDOW_SIZE)

        # Track appearances and scores for each sentence
        sentence_appearances = {i: 0 for i in range(len(sentences))}
        sentence_scores = {i: {'human_prob': 0.0, 'ai_prob': 0.0} for i in range(len(sentences))}

        # Process windows in batches
        for i in range(0, len(windows), BATCH_SIZE):
            batch_windows = windows[i:i + BATCH_SIZE]
            batch_indices = window_sentence_indices[i:i + BATCH_SIZE]

            # Tokenize and prepare input
            inputs = self.tokenizer(
                batch_windows,
                truncation=True,
                padding=True,
                max_length=MAX_LENGTH,
                return_tensors="pt"
            ).to(self.device)

            # Run inference
            with torch.no_grad():
                outputs = self.model(**inputs)
                probs = F.softmax(outputs.logits, dim=-1)

                # Process each window's predictions
                for window_idx, indices in enumerate(batch_indices):
                    center_idx = len(indices) // 2
                    center_weight = 0.7  # Center sentence gets 70% weight
                    edge_weight = 0.3 / (len(indices) - 1)  # Other sentences share 30%

                    # Apply weighted prediction to each sentence in window
                    for pos, sent_idx in enumerate(indices):
                        weight = center_weight if pos == center_idx else edge_weight
                        sentence_appearances[sent_idx] += weight
                        sentence_scores[sent_idx]['human_prob'] += weight * probs[window_idx][1].item()
                        sentence_scores[sent_idx]['ai_prob'] += weight * probs[window_idx][0].item()

            # Clean up memory
            del inputs, outputs, probs
            if torch.cuda.is_available():
                torch.cuda.empty_cache()

        # Calculate final predictions for each sentence with smoothing between adjacent sentences
        sentence_predictions = []
        for i in range(len(sentences)):
            if sentence_appearances[i] > 0:
                human_prob = sentence_scores[i]['human_prob'] / sentence_appearances[i]
                ai_prob = sentence_scores[i]['ai_prob'] / sentence_appearances[i]

                # Apply smoothing for sentences not at boundaries
                if i > 0 and i < len(sentences) - 1:
                    prev_human = sentence_scores[i-1]['human_prob'] / sentence_appearances[i-1]
                    prev_ai = sentence_scores[i-1]['ai_prob'] / sentence_appearances[i-1]
                    next_human = sentence_scores[i+1]['human_prob'] / sentence_appearances[i+1]
                    next_ai = sentence_scores[i+1]['ai_prob'] / sentence_appearances[i+1]

                    current_pred = 'human' if human_prob > ai_prob else 'ai'
                    prev_pred = 'human' if prev_human > prev_ai else 'ai'
                    next_pred = 'human' if next_human > next_ai else 'ai'

                    # Only smooth if current sentence prediction differs from neighbors
                    if current_pred != prev_pred or current_pred != next_pred:
                        smooth_factor = 0.1  # 10% smoothing factor
                        human_prob = (human_prob * (1 - smooth_factor) + 
                                    (prev_human + next_human) * smooth_factor / 2)
                        ai_prob = (ai_prob * (1 - smooth_factor) + 
                                (prev_ai + next_ai) * smooth_factor / 2)

                sentence_predictions.append({
                    'sentence': sentences[i],
                    'human_prob': human_prob,
                    'ai_prob': ai_prob,
                    'prediction': 'human' if human_prob > ai_prob else 'ai',
                    'confidence': max(human_prob, ai_prob)
                })

        # Return detailed results
        return {
            'sentence_predictions': sentence_predictions,
            'highlighted_text': self.format_predictions_html(sentence_predictions),
            'full_text': text,
            'overall_prediction': self.aggregate_predictions(sentence_predictions)
        }

    # Format predictions with color highlighting for visual assessment
    def format_predictions_html(self, sentence_predictions: List[Dict]) -> str:
        html_parts = []
        
        for pred in sentence_predictions:
            sentence = pred['sentence']
            confidence = pred['confidence']
            
            # Color coding: stronger colors for high confidence, lighter for low confidence
            if confidence >= CONFIDENCE_THRESHOLD:
                if pred['prediction'] == 'human':
                    color = "#90EE90"  # Green for human (high confidence)
                else:
                    color = "#FFB6C6"  # Pink for AI (high confidence)
            else:
                if pred['prediction'] == 'human':
                    color = "#E8F5E9"  # Light green for human (low confidence)
                else:
                    color = "#FFEBEE"  # Light pink for AI (low confidence)
                    
            html_parts.append(f'<span style="background-color: {color};">{sentence}</span>')
            
        return " ".join(html_parts)

    # Aggregate individual sentence predictions into an overall result
    def aggregate_predictions(self, predictions: List[Dict]) -> Dict:
        if not predictions:
            return {
                'prediction': 'unknown',
                'confidence': 0.0,
                'num_sentences': 0
            }

        # Calculate average probabilities across all sentences
        total_human_prob = sum(p['human_prob'] for p in predictions)
        total_ai_prob = sum(p['ai_prob'] for p in predictions)
        num_sentences = len(predictions)

        avg_human_prob = total_human_prob / num_sentences
        avg_ai_prob = total_ai_prob / num_sentences

        return {
            'prediction': 'human' if avg_human_prob > avg_ai_prob else 'ai',
            'confidence': max(avg_human_prob, avg_ai_prob),
            'num_sentences': num_sentences
        }

# MAIN ANALYSIS FUNCTION
# Brings everything together to analyze text based on selected mode
def analyze_text(text: str, mode: str, classifier: TextClassifier) -> tuple:
    start_time = time.time()
    
    word_count = len(text.split())
    
    # Auto-switch to quick mode for short texts
    original_mode = mode
    if word_count < 200 and mode == "detailed":
        mode = "quick"
    
    if mode == "quick":
        # Perform quick analysis
        result = classifier.quick_scan(text)
        
        quick_analysis = f"""
        PREDICTION: {result['prediction'].upper()}
        Confidence: {result['confidence']*100:.1f}%
        Windows analyzed: {result['num_windows']}
        """
        
        # Notify if automatically switched from detailed to quick mode
        if original_mode == "detailed":
            quick_analysis += f"\n\nNote: Switched to quick mode because text contains only {word_count} words. Minimum 200 words required for detailed analysis."
        
        execution_time = (time.time() - start_time) * 1000
        
        return (
            text,  # Original text (no highlighting)
            "Quick scan mode - no sentence-level analysis available",
            quick_analysis
        )
    else:
        # Perform detailed analysis
        analysis = classifier.detailed_scan(text)
        
        # Format sentence-by-sentence analysis text
        detailed_analysis = []
        for pred in analysis['sentence_predictions']:
            confidence = pred['confidence'] * 100
            detailed_analysis.append(f"Sentence: {pred['sentence']}")
            detailed_analysis.append(f"Prediction: {pred['prediction'].upper()}")
            detailed_analysis.append(f"Confidence: {confidence:.1f}%")
            detailed_analysis.append("-" * 50)
        
        # Format overall result summary
        final_pred = analysis['overall_prediction']
        overall_result = f"""
        FINAL PREDICTION: {final_pred['prediction'].upper()}
        Overall confidence: {final_pred['confidence']*100:.1f}%
        Number of sentences analyzed: {final_pred['num_sentences']}
        """
        
        execution_time = (time.time() - start_time) * 1000
        
        return (
            analysis['highlighted_text'],  # HTML-highlighted text
            "\n".join(detailed_analysis),  # Detailed sentence analysis
            overall_result  # Overall summary
        )

# Initialize the classifier
classifier = TextClassifier()

# GRADIO USER INTERFACE
demo = gr.Interface(
    fn=lambda text, mode: analyze_text(text, mode, classifier),
    inputs=[
        gr.Textbox(
            lines=8,
            placeholder="Enter text to analyze...",
            label="Input Text"
        ),
        gr.Radio(
            choices=["quick", "detailed"],
            value="quick",
            label="Analysis Mode",
            info="Quick mode for faster analysis, Detailed mode for sentence-level analysis"
        )
    ],
    outputs=[
        gr.HTML(label="Highlighted Analysis"),  # Shows color-coded result
        gr.Textbox(label="Sentence-by-Sentence Analysis", lines=10),  # Detailed breakdown
        gr.Textbox(label="Overall Result", lines=4)  # Summary results
    ],
    title="AI Text Detector",
    description="Analyze text to detect if it was written by a human or AI. Choose between quick scan and detailed sentence-level analysis. 200+ words suggested for accurate predictions.",
    api_name="predict",
    flagging_mode="never"
)

# FastAPI configuration
app = demo.app

# Add CORS middleware to allow cross-origin requests
app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["GET", "POST", "OPTIONS"],
    allow_headers=["*"],
)

# Start the server when run directly
if __name__ == "__main__":
    demo.queue()  # Enable request queuing
    demo.launch(
        server_name="0.0.0.0",  # Listen on all interfaces
        server_port=7860,  # Default Gradio port
        share=True  # Generate public URL
    )