File size: 11,906 Bytes
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
379daf7
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a1ce4b0
0f03dd5
 
 
1afd410
c0eef1e
1afd410
c0eef1e
 
 
 
 
1afd410
c0eef1e
 
 
 
 
0f03dd5
 
 
 
 
c0eef1e
0f03dd5
1afd410
 
 
 
0f03dd5
 
 
1afd410
0f03dd5
 
c0eef1e
 
 
5ad87e4
1afd410
 
 
c0eef1e
1afd410
 
c0eef1e
 
1afd410
0f03dd5
1afd410
 
 
 
 
 
0f03dd5
1afd410
0f03dd5
 
1afd410
c0eef1e
0f03dd5
1afd410
 
0f03dd5
 
 
a1ce4b0
0f03dd5
 
1afd410
0f03dd5
1afd410
 
 
a1ce4b0
1afd410
 
0f03dd5
 
 
 
1afd410
8663fbd
c0eef1e
1afd410
0f03dd5
1afd410
0f03dd5
1afd410
 
 
 
 
0f03dd5
 
 
1afd410
0f03dd5
a1ce4b0
0f03dd5
 
 
1afd410
8663fbd
c0eef1e
1afd410
0f03dd5
1afd410
0f03dd5
1afd410
0f03dd5
 
a1ce4b0
 
1afd410
a1ce4b0
 
 
 
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
1afd410
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1afd410
0f03dd5
 
1afd410
 
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a1ce4b0
 
 
 
 
 
 
0f03dd5
a1ce4b0
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1afd410
 
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import json
import numpy as np
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForQuestionAnswering, pipeline
import torch
from sklearn.metrics import f1_score
import re
from collections import Counter
import string
from huggingface_hub import login
import gradio as gr
import pandas as pd
from datetime import datetime

def normalize_answer(s):
    """Normalize answer for evaluation"""
    def remove_articles(text):
        return re.sub(r'\b(a|an|the)\b', ' ', text)
    
    def white_space_fix(text):
        return ' '.join(text.split())
    
    def remove_punc(text):
        exclude = set(string.punctuation)
        return ''.join(ch for ch in text if ch not in exclude)
    
    def lower(text):
        return text.lower()
    
    return white_space_fix(remove_articles(remove_punc(lower(s))))

def f1_score_qa(prediction, ground_truth):
    """Calculate F1 score for QA"""
    prediction_tokens = normalize_answer(prediction).split()
    ground_truth_tokens = normalize_answer(ground_truth).split()
    
    if len(prediction_tokens) == 0 or len(ground_truth_tokens) == 0:
        return int(prediction_tokens == ground_truth_tokens)
    
    common = Counter(prediction_tokens) & Counter(ground_truth_tokens)
    num_same = sum(common.values())
    
    if num_same == 0:
        return 0
    
    precision = 1.0 * num_same / len(prediction_tokens)
    recall = 1.0 * num_same / len(ground_truth_tokens)
    f1 = (2 * precision * recall) / (precision + recall)
    return f1

def exact_match_score(prediction, ground_truth):
    """Calculate exact match score"""
    return normalize_answer(prediction) == normalize_answer(ground_truth)

def evaluate_model():
    # Authenticate with Hugging Face using the token
    hf_token = os.getenv("EVAL_TOKEN")
    if hf_token:
        try:
            login(token=hf_token)
            print("βœ“ Authenticated with Hugging Face")
        except Exception as e:
            print(f"⚠ Warning: Could not authenticate with HF token: {e}")
    else:
        print("⚠ Warning: EVAL_TOKEN not found in environment variables")
    
    print("Loading model and tokenizer...")
    model_name = "AvocadoMuffin/roberta-cuad-qa-v2"
    
    try:
        tokenizer = AutoTokenizer.from_pretrained(model_name, token=hf_token)
        model = AutoModelForQuestionAnswering.from_pretrained(model_name, token=hf_token)
        qa_pipeline = pipeline("question-answering", model=model, tokenizer=tokenizer)
        print("βœ“ Model loaded successfully")
        return qa_pipeline, hf_token
    except Exception as e:
        print(f"βœ— Error loading model: {e}")
        return None, None

def run_evaluation(num_samples, progress=gr.Progress()):
    """Run evaluation and return results for Gradio interface"""
    
    # Load model
    qa_pipeline, hf_token = evaluate_model()
    if qa_pipeline is None:
        return "❌ Failed to load model", pd.DataFrame(), None
    
    progress(0.1, desc="Loading CUAD dataset...")
    
    # Load dataset - use QA format version (JSON, no PDFs)
    try:
        # Try the QA-specific version first (much faster, JSON format)
        dataset = load_dataset("theatticusproject/cuad-qa", trust_remote_code=True, token=hf_token)
        test_data = dataset["test"]
        print(f"βœ“ Loaded CUAD-QA dataset with {len(test_data)} samples")
    except Exception as e:
        try:
            # Fallback to original but limit to avoid PDF downloads
            dataset = load_dataset("cuad", split="test[:1000]", trust_remote_code=True, token=hf_token)
            test_data = dataset
            print(f"βœ“ Loaded CUAD dataset with {len(test_data)} samples")
        except Exception as e2:
            return f"❌ Error loading dataset: {e2}", pd.DataFrame(), None
    
    # Limit samples
    num_samples = min(num_samples, len(test_data))
    test_subset = test_data.select(range(num_samples))
    
    progress(0.2, desc=f"Starting evaluation on {num_samples} samples...")
    
    # Initialize metrics
    exact_matches = []
    f1_scores = []
    predictions = []
    
    # Run evaluation
    for i, example in enumerate(test_subset):
        progress((0.2 + 0.7 * i / num_samples), desc=f"Processing sample {i+1}/{num_samples}")
        
        try:
            context = example["context"]
            question = example["question"]
            answers = example["answers"]
            
            # Get model prediction
            result = qa_pipeline(question=question, context=context)
            predicted_answer = result["answer"]
            
            # Get ground truth answers
            if answers["text"] and len(answers["text"]) > 0:
                ground_truth = answers["text"][0] if isinstance(answers["text"], list) else answers["text"]
            else:
                ground_truth = ""
            
            # Calculate metrics
            em = exact_match_score(predicted_answer, ground_truth)
            f1 = f1_score_qa(predicted_answer, ground_truth)
            
            exact_matches.append(em)
            f1_scores.append(f1)
            
            predictions.append({
                "Sample_ID": i+1,
                "Question": question[:100] + "..." if len(question) > 100 else question,
                "Predicted_Answer": predicted_answer,
                "Ground_Truth": ground_truth,
                "Exact_Match": em,
                "F1_Score": round(f1, 3),
                "Confidence": round(result["score"], 3)
            })
            
        except Exception as e:
            print(f"Error processing sample {i}: {e}")
            continue
    
    progress(0.9, desc="Calculating final metrics...")
    
    # Calculate final metrics
    if len(exact_matches) == 0:
        return "❌ No samples were successfully processed", pd.DataFrame(), None
    
    avg_exact_match = np.mean(exact_matches) * 100
    avg_f1_score = np.mean(f1_scores) * 100
    
    # Create results summary
    results_summary = f"""
# πŸ“Š CUAD Model Evaluation Results
## 🎯 Overall Performance
- **Model**: AvocadoMuffin/roberta-cuad-qa-v3
- **Dataset**: CUAD (Contract Understanding Atticus Dataset)
- **Samples Evaluated**: {len(exact_matches)}
- **Evaluation Date**: {datetime.now().strftime("%Y-%m-%d %H:%M:%S")}
## πŸ“ˆ Metrics
- **Exact Match Score**: {avg_exact_match:.2f}%
- **F1 Score**: {avg_f1_score:.2f}%
## πŸ” Performance Analysis
- **High Confidence Predictions**: {len([p for p in predictions if p['Confidence'] > 0.8])} ({len([p for p in predictions if p['Confidence'] > 0.8])/len(predictions)*100:.1f}%)
- **Perfect Matches**: {len([p for p in predictions if p['Exact_Match'] == 1])} ({len([p for p in predictions if p['Exact_Match'] == 1])/len(predictions)*100:.1f}%)
- **High F1 Scores (>0.8)**: {len([p for p in predictions if p['F1_Score'] > 0.8])} ({len([p for p in predictions if p['F1_Score'] > 0.8])/len(predictions)*100:.1f}%)
"""
    
    # Create detailed results DataFrame
    df = pd.DataFrame(predictions)
    
    # Save results to file
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    results_file = f"cuad_evaluation_results_{timestamp}.json"
    
    detailed_results = {
        "model_name": "AvocadoMuffin/roberta-cuad-qa-v3",
        "dataset": "cuad",
        "num_samples": len(exact_matches),
        "exact_match_score": avg_exact_match,
        "f1_score": avg_f1_score,
        "evaluation_date": datetime.now().isoformat(),
        "predictions": predictions
    }
    
    try:
        with open(results_file, "w") as f:
            json.dump(detailed_results, f, indent=2)
        print(f"βœ“ Results saved to {results_file}")
    except Exception as e:
        print(f"⚠ Warning: Could not save results file: {e}")
        results_file = None
    
    progress(1.0, desc="βœ… Evaluation completed!")
    
    return results_summary, df, results_file

def create_gradio_interface():
    """Create Gradio interface for CUAD evaluation"""
    
    with gr.Blocks(title="CUAD Model Evaluator", theme=gr.themes.Soft()) as demo:
        gr.HTML("""
        <div style="text-align: center; padding: 20px;">
            <h1>πŸ›οΈ CUAD Model Evaluation Dashboard</h1>
            <p>Evaluate your CUAD (Contract Understanding Atticus Dataset) Question Answering model</p>
            <p><strong>Model:</strong> AvocadoMuffin/roberta-cuad-qa-v2</p>
        </div>
        """)
        
        with gr.Row():
            with gr.Column(scale=1):
                gr.HTML("<h3>βš™οΈ Evaluation Settings</h3>")
                
                num_samples = gr.Slider(
                    minimum=10,
                    maximum=500,
                    value=100,
                    step=10,
                    label="Number of samples to evaluate",
                    info="Choose between 10-500 samples (more samples = more accurate but slower)"
                )
                
                evaluate_btn = gr.Button(
                    "πŸš€ Start Evaluation", 
                    variant="primary",
                    size="lg"
                )
                
                gr.HTML("""
                <div style="margin-top: 20px; padding: 15px; background-color: #f0f0f0; border-radius: 8px;">
                    <h4>πŸ“‹ What this evaluates:</h4>
                    <ul>
                        <li><strong>Exact Match</strong>: Percentage of perfect predictions</li>
                        <li><strong>F1 Score</strong>: Token-level overlap between prediction and ground truth</li>
                        <li><strong>Confidence</strong>: Model's confidence in its predictions</li>
                    </ul>
                </div>
                """)
            
            with gr.Column(scale=2):
                gr.HTML("<h3>πŸ“Š Results</h3>")
                
                results_summary = gr.Markdown(
                    value="Click 'πŸš€ Start Evaluation' to begin...",
                    label="Evaluation Summary"
                )
        
        gr.HTML("<hr>")
        
        with gr.Row():
            gr.HTML("<h3>πŸ“‹ Detailed Results</h3>")
        
        with gr.Row():
            detailed_results = gr.Dataframe(
                label="Sample-by-Sample Results",
                interactive=False,
                wrap=True
            )
        
        with gr.Row():
            download_file = gr.File(
                label="πŸ“₯ Download Complete Results (JSON)",
                visible=False
            )
        
        # Event handlers
        def handle_evaluation(num_samples):
            summary, df, file_path = run_evaluation(num_samples)
            if file_path and os.path.exists(file_path):
                return summary, df, gr.update(visible=True, value=file_path)
            else:
                return summary, df, gr.update(visible=False)
        
        evaluate_btn.click(
            fn=handle_evaluation,
            inputs=[num_samples],
            outputs=[results_summary, detailed_results, download_file],
            show_progress=True
        )
        
        # Footer
        gr.HTML("""
        <div style="text-align: center; margin-top: 30px; padding: 20px; color: #666;">
            <p>πŸ€– Powered by Hugging Face Transformers & Gradio</p>
            <p>πŸ“š CUAD Dataset by The Atticus Project</p>
        </div>
        """)
    
    return demo
    
if __name__ == "__main__":
    print("CUAD Model Evaluation with Gradio Interface")
    print("=" * 50)
    
    # Check if CUDA is available
    if torch.cuda.is_available():
        print(f"βœ“ CUDA available: {torch.cuda.get_device_name(0)}")
    else:
        print("! Running on CPU")
    
    # Create and launch Gradio interface
    demo = create_gradio_interface()
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=True,
        debug=True
    )