File size: 17,418 Bytes
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f6c6d61
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
4c1232a
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a1ce4b0
0f03dd5
 
 
f6c6d61
c0eef1e
 
 
 
 
 
 
 
 
 
 
0f03dd5
 
 
 
 
c0eef1e
0f03dd5
f6c6d61
 
 
 
 
 
 
0f03dd5
 
 
f6c6d61
0f03dd5
 
c0eef1e
 
 
5ad87e4
f6c6d61
 
 
 
 
 
 
 
c0eef1e
f6c6d61
 
c0eef1e
 
f6c6d61
0f03dd5
f6c6d61
 
 
 
 
 
 
 
 
0f03dd5
f6c6d61
0f03dd5
 
f6c6d61
 
 
c0eef1e
f6c6d61
 
0f03dd5
f6c6d61
0f03dd5
 
 
a1ce4b0
0f03dd5
 
f6c6d61
0f03dd5
f6c6d61
 
 
a1ce4b0
f6c6d61
 
 
 
 
 
 
 
 
 
 
 
 
0f03dd5
 
 
 
f6c6d61
 
8663fbd
c0eef1e
f6c6d61
 
0f03dd5
f6c6d61
 
 
 
 
 
0f03dd5
f6c6d61
 
 
 
 
 
 
 
 
 
 
 
0f03dd5
 
 
f6c6d61
0f03dd5
a1ce4b0
0f03dd5
 
 
f6c6d61
8663fbd
c0eef1e
f6c6d61
 
 
 
0f03dd5
f6c6d61
 
0f03dd5
f6c6d61
0f03dd5
 
a1ce4b0
 
f6c6d61
a1ce4b0
 
 
 
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
f6c6d61
 
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f6c6d61
0f03dd5
f6c6d61
 
0f03dd5
f6c6d61
0f03dd5
f6c6d61
0f03dd5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a1ce4b0
 
 
 
 
 
 
0f03dd5
a1ce4b0
0f03dd5
 
 
 
 
 
 
 
 
 
f6c6d61
0f03dd5
 
 
 
 
 
f6c6d61
 
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
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
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 has_answer(answers):
    """Check if the question has any valid answers"""
    if not answers or not answers.get("text"):
        return False
    
    answer_texts = answers["text"] if isinstance(answers["text"], list) else [answers["text"]]
    return any(text.strip() for text in answer_texts)

def get_top_k_predictions(qa_pipeline, question, context, k=3):
    """Get top-k predictions from the model"""
    # Get raw model outputs
    inputs = qa_pipeline.tokenizer(question, context, return_tensors="pt", truncation=True, max_length=512)
    
    with torch.no_grad():
        outputs = qa_pipeline.model(**inputs)
        start_logits = outputs.start_logits
        end_logits = outputs.end_logits
    
    # Get top-k start and end positions
    start_scores, start_indices = torch.topk(start_logits.flatten(), k)
    end_scores, end_indices = torch.topk(end_logits.flatten(), k)
    
    predictions = []
    
    # Generate all combinations of start and end positions
    for start_idx in start_indices:
        for end_idx in end_indices:
            if start_idx <= end_idx:  # Valid span
                # Convert to answer text
                input_ids = inputs["input_ids"][0]
                answer_tokens = input_ids[start_idx:end_idx + 1]
                answer_text = qa_pipeline.tokenizer.decode(answer_tokens, skip_special_tokens=True)
                
                # Calculate combined score
                start_score = start_logits[0][start_idx].item()
                end_score = end_logits[0][end_idx].item()
                combined_score = start_score + end_score
                
                predictions.append({
                    "answer": answer_text,
                    "score": combined_score,
                    "start": start_idx.item(),
                    "end": end_idx.item()
                })
    
    # Sort by score and return top-k unique answers
    predictions.sort(key=lambda x: x["score"], reverse=True)
    unique_answers = []
    seen_answers = set()
    
    for pred in predictions:
        normalized_answer = normalize_answer(pred["answer"])
        if normalized_answer not in seen_answers and len(unique_answers) < k:
            unique_answers.append(pred)
            seen_answers.add(normalized_answer)
    
    return unique_answers

def calculate_top_k_has_ans_f1(predictions, ground_truths, k=1):
    """Calculate Top-K Has Answer F1 score"""
    f1_scores = []
    
    for preds, gt in zip(predictions, ground_truths):
        if not has_answer(gt):
            continue  # Skip questions without answers
        
        # Get ground truth text
        gt_text = gt["text"][0] if isinstance(gt["text"], list) else gt["text"]
        
        # Calculate F1 for top-k predictions
        max_f1 = 0
        for i in range(min(k, len(preds))):
            pred_text = preds[i]["answer"]
            f1 = f1_score_qa(pred_text, gt_text)
            max_f1 = max(max_f1, f1)
        
        f1_scores.append(max_f1)
    
    return np.mean(f1_scores) if f1_scores else 0

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
    try:
        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:
            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 storage for predictions and ground truths
    all_top_k_predictions = []
    all_ground_truths = []
    all_has_answer_flags = []
    
    # Storage for detailed results
    detailed_results = []
    
    # Run evaluation
    for i, example in enumerate(test_subset):
        progress((0.2 + 0.6 * i / num_samples), desc=f"Processing sample {i+1}/{num_samples}")
        
        try:
            context = example["context"]
            question = example["question"]
            answers = example["answers"]
            
            # Check if question has answers
            has_ans = has_answer(answers)
            all_has_answer_flags.append(has_ans)
            all_ground_truths.append(answers)
            
            # Get top-3 predictions
            top_k_preds = get_top_k_predictions(qa_pipeline, question, context, k=3)
            all_top_k_predictions.append(top_k_preds)
            
            # Get ground truth for display
            if has_ans:
                ground_truth = answers["text"][0] if isinstance(answers["text"], list) else answers["text"]
            else:
                ground_truth = "[No Answer]"
            
            # Calculate metrics for this sample
            if has_ans and top_k_preds:
                top1_f1 = f1_score_qa(top_k_preds[0]["answer"], ground_truth)
                top3_f1 = max([f1_score_qa(pred["answer"], ground_truth) for pred in top_k_preds[:3]])
                em = exact_match_score(top_k_preds[0]["answer"], ground_truth)
            else:
                top1_f1 = 0
                top3_f1 = 0
                em = 0
            
            detailed_results.append({
                "Sample_ID": i+1,
                "Question": question[:100] + "..." if len(question) > 100 else question,
                "Has_Answer": has_ans,
                "Top1_Prediction": top_k_preds[0]["answer"] if top_k_preds else "[No Prediction]",
                "Top3_Predictions": " | ".join([p["answer"] for p in top_k_preds[:3]]),
                "Ground_Truth": ground_truth,
                "Top1_F1": round(top1_f1, 3),
                "Top3_F1": round(top3_f1, 3),
                "Exact_Match": em,
                "Top1_Confidence": round(top_k_preds[0]["score"], 3) if top_k_preds else 0
            })
            
        except Exception as e:
            print(f"Error processing sample {i}: {e}")
            continue
    
    progress(0.8, desc="Calculating final metrics...")
    
    # Filter for questions with answers only
    has_ans_predictions = [pred for pred, has_ans in zip(all_top_k_predictions, all_has_answer_flags) if has_ans]
    has_ans_ground_truths = [gt for gt, has_ans in zip(all_ground_truths, all_has_answer_flags) if has_ans]
    
    if len(has_ans_predictions) == 0:
        return "❌ No samples with answers were found", pd.DataFrame(), None
    
    # Calculate Top-K Has Answer F1 scores
    top1_has_ans_f1 = calculate_top_k_has_ans_f1(has_ans_predictions, has_ans_ground_truths, k=1) * 100
    top3_has_ans_f1 = calculate_top_k_has_ans_f1(has_ans_predictions, has_ans_ground_truths, k=3) * 100
    
    # Calculate overall metrics
    total_samples = len(detailed_results)
    has_answer_samples = len(has_ans_predictions)
    avg_exact_match = np.mean([r["Exact_Match"] for r in detailed_results]) * 100
    avg_top1_f1 = np.mean([r["Top1_F1"] for r in detailed_results if r["Has_Answer"]]) * 100
    avg_top3_f1 = np.mean([r["Top3_F1"] for r in detailed_results if r["Has_Answer"]]) * 100
    
    # Create results summary
    results_summary = f"""
# πŸ“Š CUAD Model Evaluation Results

## 🎯 Model Performance
- **Model**: AvocadoMuffin/roberta-cuad-qa-v3
- **Dataset**: CUAD (Contract Understanding Atticus Dataset)
- **Total Samples**: {total_samples}
- **Samples with Answers**: {has_answer_samples}
- **Evaluation Date**: {datetime.now().strftime("%Y-%m-%d %H:%M:%S")}

## πŸ“ˆ Key Metrics (Industry Standard)
- **Top 1 Has Ans F1**: {top1_has_ans_f1:.2f}%
- **Top 3 Has Ans F1**: {top3_has_ans_f1:.2f}%

## πŸ“‹ Additional Metrics
- **Exact Match Score**: {avg_exact_match:.2f}%
- **Average Top-1 F1**: {avg_top1_f1:.2f}%
- **Average Top-3 F1**: {avg_top3_f1:.2f}%

## πŸ” Performance Breakdown
- **High Confidence Predictions (>0.8)**: {len([r for r in detailed_results if r['Top1_Confidence'] > 0.8])} ({len([r for r in detailed_results if r['Top1_Confidence'] > 0.8])/total_samples*100:.1f}%)
- **Perfect Matches**: {len([r for r in detailed_results if r['Exact_Match'] == 1])} ({len([r for r in detailed_results if r['Exact_Match'] == 1])/total_samples*100:.1f}%)
- **High F1 Scores (>0.8)**: {len([r for r in detailed_results if r['Top1_F1'] > 0.8])} ({len([r for r in detailed_results if r['Top1_F1'] > 0.8])/has_answer_samples*100:.1f}%)

## πŸ“Š Comparison with Benchmarks
Your model's **Top 1 Has Ans F1** of {top1_has_ans_f1:.2f}% can be compared to:
- gustavhartz/roberta-base-cuad-finetuned: 85.68%
- Rakib/roberta-base-on-cuad: 81.26%
"""
    
    # Create detailed results DataFrame
    df = pd.DataFrame(detailed_results)
    
    # Save results to file
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    results_file = f"cuad_evaluation_results_{timestamp}.json"
    
    complete_results = {
        "model_name": "AvocadoMuffin/roberta-cuad-qa-v3",
        "dataset": "cuad",
        "total_samples": total_samples,
        "has_answer_samples": has_answer_samples,
        "top1_has_ans_f1": top1_has_ans_f1,
        "top3_has_ans_f1": top3_has_ans_f1,
        "exact_match_score": avg_exact_match,
        "avg_top1_f1": avg_top1_f1,
        "avg_top3_f1": avg_top3_f1,
        "evaluation_date": datetime.now().isoformat(),
        "detailed_results": detailed_results
    }
    
    try:
        with open(results_file, "w") as f:
            json.dump(complete_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-v3</p>
            <p><em>Now with industry-standard Top-K Has Answer F1 metrics!</em></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“‹ Evaluation Metrics:</h4>
                    <ul>
                        <li><strong>Top 1 Has Ans F1</strong>: F1 score for single best answer (industry standard)</li>
                        <li><strong>Top 3 Has Ans F1</strong>: F1 score allowing up to 3 predictions</li>
                        <li><strong>Exact Match</strong>: Percentage of perfect predictions</li>
                        <li><strong>Confidence</strong>: Model's confidence in predictions</li>
                    </ul>
                    <p><em>Note: "Has Ans" metrics only consider questions that have valid answers.</em></p>
                </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>
            <p>πŸ“Š Now with industry-standard Top-K Has Answer F1 metrics</p>
        </div>
        """)
    
    return demo
    
if __name__ == "__main__":
    print("CUAD Model Evaluation with Top-K Has Answer F1 Metrics")
    print("=" * 60)
    
    # 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
    )