File size: 19,765 Bytes
ad39573
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
import os
import json
import pandas as pd
import glob
import gradio as gr

# Cache for loaded data
data_cache = {}

# Load data functions with caching
def load_jsonl(file_path):
    """Load a JSONL file into a pandas DataFrame with caching."""
    if file_path in data_cache:
        return data_cache[file_path]
    
    if not os.path.exists(file_path):
        return pd.DataFrame()
    
    try:
        df = pd.read_json(file_path, lines=True)
        data_cache[file_path] = df
        return df
    except Exception as e:
        print(f"Error loading {file_path}: {e}")
        return pd.DataFrame()

def get_available_benchmarks():
    """Get list of available benchmarks in data directory."""
    return [dir_name for dir_name in os.listdir("data") 
            if os.path.isdir(os.path.join("data", dir_name))]

def get_categories(benchmark):
    """Get list of categories for a given benchmark."""
    questions = load_jsonl(f"data/{benchmark}/question.jsonl")
    if questions.empty:
        return []
    return sorted(questions['category'].unique().tolist())

def get_languages(benchmark):
    """Get list of languages available in the benchmark."""
    questions = load_jsonl(f"data/{benchmark}/question.jsonl")
    if questions.empty or 'language' not in questions.columns:
        return ["English"]  # Default if no language column
    
    return sorted(questions['language'].unique().tolist())

def get_judges(benchmark):
    """Get list of available judges for a benchmark."""
    judgment_dir = f"data/{benchmark}/model_judgment"
    if not os.path.exists(judgment_dir):
        return []
    return [dir_name for dir_name in os.listdir(judgment_dir)
            if os.path.isdir(os.path.join(judgment_dir, dir_name))]

def get_models(benchmark, judge):
    """Get list of models that have judgments by the specified judge."""
    if not judge:
        return []
    
    judgment_dir = f"data/{benchmark}/model_judgment/{judge}"
    if not os.path.exists(judgment_dir):
        return []
    
    return [os.path.splitext(os.path.basename(file))[0] 
            for file in glob.glob(f"{judgment_dir}/*.jsonl")]

def get_questions(benchmark, category=None, language=None):
    """Get questions with category and language filters if provided."""
    questions = load_jsonl(f"data/{benchmark}/question.jsonl")
    if questions.empty:
        return []
    
    # Apply category filter if provided
    if category and category != "All":
        questions = questions[questions['category'] == category]
    
    # Apply language filter if provided and column exists
    if language and language != "All" and 'language' in questions.columns:
        questions = questions[questions['language'] == language]
    
    # Create list of question previews with their UIDs
    question_previews = [(row['uid'], row['prompt'][:100] + "..." if len(row['prompt']) > 100 else row['prompt']) 
                        for _, row in questions.iterrows()]
    
    return question_previews

def get_model_answer(benchmark, model, uid):
    """Get a model's answer for a specific question."""
    model_answers = load_jsonl(f"data/{benchmark}/model_answer/{model}.jsonl")
    if model_answers.empty:
        return "No answer found"
    
    answer = model_answers[model_answers['uid'] == uid]
    if answer.empty:
        return "No answer found"
    
    # Extract the actual answer from the messages
    try:
        messages = answer.iloc[0]['messages']
        if len(messages) < 2:
            return "No answer found"
        
        # The assistant's message should be the second one
        assistant_msg = messages[1]
        if 'role' in assistant_msg and assistant_msg['role'] == 'assistant':
            content = assistant_msg['content']
            
            # Handle different content formats
            if isinstance(content, dict) and 'answer' in content:
                return content['answer']
            elif isinstance(content, str):
                return content
            else:
                return str(content)
        else:
            return "Invalid message format"
    except Exception as e:
        return f"Error extracting answer: {str(e)}"

def get_judgment(benchmark, judge, model, uid):
    """Get judgment for a specific model and question."""
    judgments = load_jsonl(f"data/{benchmark}/model_judgment/{judge}/{model}.jsonl")
    if judgments.empty:
        return None, None
    
    judgment = judgments[judgments['uid'] == uid]
    if judgment.empty:
        return None, None
    
    games = judgment.iloc[0]['games']
    if len(games) < 2:
        return games[0] if games else None, None
    
    return games[0], games[1]  # First game, second game

def format_judgment(game):
    """Format judgment for display."""
    if not game:
        return "No judgment available"
    
    score = game.get('score', 'No score')
    
    # Try to get judgment text
    judgment = game.get('judgment', {})
    if isinstance(judgment, dict) and 'answer' in judgment:
        judgment_text = judgment['answer']
    else:
        judgment_text = str(judgment)
    
    return f"### Score: {score}\n\n{judgment_text}"

# Gradio interface functions
def update_categories(benchmark):
    """Update category dropdown based on selected benchmark."""
    categories = ["All"] + get_categories(benchmark)
    return gr.Dropdown(choices=categories, value="All")

def update_languages(benchmark):
    """Update language dropdown based on selected benchmark."""
    languages = ["All"] + get_languages(benchmark)
    default = "English" if "English" in languages else languages[0]
    return gr.Dropdown(choices=languages, value=default)

def update_judges(benchmark):
    """Update judge dropdown based on selected benchmark."""
    judges = get_judges(benchmark)
    default = judges[0] if judges else None
    return gr.Dropdown(choices=judges, value=default)

def update_models(benchmark, judge):
    """Update model dropdown based on selected benchmark and judge."""
    models = get_models(benchmark, judge)
    default = models[0] if models else None
    return gr.Dropdown(choices=models, value=default)

def update_questions(benchmark, category, language):
    """Update question dropdown based on selected benchmark, category and language."""
    question_list = get_questions(benchmark, category, language)
    if not question_list:
        return gr.Dropdown(choices=[], value=None), {}
    
    # Create a dictionary mapping previews to UIDs to ensure we can look up UIDs from previews
    question_dict = {q[1]: q[0] for q in question_list}
    question_options = list(question_dict.keys())
    
    default = question_options[0] if question_options else None
    return gr.Dropdown(choices=question_options, value=default), question_dict

def display_content(benchmark, category, language, judge, model, question, question_dict):
    """Display the question, answers, and judgments."""
    if not question or not question_dict or question not in question_dict:
        return "No question selected", "No baseline answer", "No model answer", "No judgment", "No judgment"
    
    uid = question_dict[question]
    
    # Load the question text
    questions_df = load_jsonl(f"data/{benchmark}/question.jsonl")
    question_row = questions_df[questions_df['uid'] == uid]
    if question_row.empty:
        return "Question not found", "No baseline answer", "No model answer", "No judgment", "No judgment"
    
    question_text = question_row.iloc[0]['prompt']
    
    # Load judgments and identify baseline model
    judgments = load_jsonl(f"data/{benchmark}/model_judgment/{judge}/{model}.jsonl")
    judgment_row = judgments[judgments['uid'] == uid]
    
    if judgment_row.empty:
        return question_text, "No baseline answer", "No model answer", "No judgment", "No judgment"
    
    baseline_model = judgment_row.iloc[0]['baseline']
    
    # Get answers
    baseline_answer = get_model_answer(benchmark, baseline_model, uid)
    model_answer = get_model_answer(benchmark, model, uid)
    
    # Get judgments
    game1, game2 = get_judgment(benchmark, judge, model, uid)
    
    judgment1 = format_judgment(game1)
    judgment2 = format_judgment(game2)
    
    return question_text, baseline_answer, model_answer, judgment1, judgment2

# Initialize app components based on selected benchmark
def init_app(benchmark):
    categories = ["All"] + get_categories(benchmark)
    default_category = "All"
    
    languages = ["All"] + get_languages(benchmark)
    default_language = "English" if "English" in languages else languages[0]
    
    judges = get_judges(benchmark)
    default_judge = judges[0] if judges else None
    
    models = get_models(benchmark, default_judge) if default_judge else []
    default_model = models[0] if models else None
    
    question_list = get_questions(benchmark, default_category, default_language)
    question_dict = {q[1]: q[0] for q in question_list}
    question_options = list(question_dict.keys())
    default_question = question_options[0] if question_options else None
    
    # Get initial display content
    if default_question and default_model and default_judge:
        question_text, baseline_ans, model_ans, judgment1, judgment2 = display_content(
            benchmark, default_category, default_language, default_judge, default_model, default_question, question_dict
        )
    else:
        question_text = "No question available"
        baseline_ans = "No baseline answer"
        model_ans = "No model answer"
        judgment1 = "No judgment"
        judgment2 = "No judgment"
    
    return (
        gr.Dropdown(choices=categories, value=default_category),
        gr.Dropdown(choices=languages, value=default_language),
        gr.Dropdown(choices=judges, value=default_judge),
        gr.Dropdown(choices=models, value=default_model),
        gr.Dropdown(choices=question_options, value=default_question),
        question_dict,
        question_text,
        baseline_ans, model_ans,
        judgment1, judgment2
    )

# Function to go to the next question
def next_question(benchmark, category, language, current_question, question_dict):
    question_list = get_questions(benchmark, category, language)
    previews = [q[1] for q in question_list]
    
    if current_question not in previews:
        return gr.Dropdown(value=previews[0] if previews else None)
            
    current_idx = previews.index(current_question)
    next_idx = (current_idx + 1) % len(previews)
    return gr.Dropdown(value=previews[next_idx])

# Create Gradio app
def create_app():
    benchmarks = get_available_benchmarks()
    default_benchmark = "arena-hard-v2.0" if "arena-hard-v2.0" in benchmarks else benchmarks[0]
    
    # Initialize data for the default benchmark
    init_data = init_app(default_benchmark)
    
    with gr.Blocks() as app:
        gr.Markdown(
            '''# Arena-Hard-Auto Benchmark Viewer
            
            Arena-Hard-Auto is an automatic evaluation tool for instruction-tuned LLMs. It has the highest correlation and separability to LMArena (Chatbot Arena) among popular open-ended LLM benchmarks. If you are curious to see how well your model might perform on LMArena before deploying, we recommend trying Arena-Hard-Auto's newest evaluation set, **Arena-Hard-v2.0-Preview**.
            
            **Repo:** https://github.com/lmarena/arena-hard-auto
            
            **Paper:** https://arxiv.org/abs/2406.11939
            '''
        )
        
        with gr.Row():
            with gr.Column():
                benchmark_dropdown = gr.Dropdown(
                    choices=benchmarks,
                    value=default_benchmark,
                    label="Benchmark"
                )
                
                category_dropdown = gr.Dropdown(
                    choices=init_data[0].choices,
                    value=init_data[0].value,
                    label="Category"
                )
                
                language_dropdown = gr.Dropdown(
                    choices=init_data[1].choices,
                    value=init_data[1].value,
                    label="Language"
                )
            
            with gr.Column():
                judge_dropdown = gr.Dropdown(
                    choices=init_data[2].choices,
                    value=init_data[2].value,
                    label="Judge Model"
                )
                
                model_dropdown = gr.Dropdown(
                    label="Model to Evaluate",
                    choices=init_data[3].choices,
                    value=init_data[3].value,
                )
        
        question_dict = gr.State(init_data[5])
        question_dropdown = gr.Dropdown(
            choices=init_data[4].choices,
            value=init_data[4].value,
            label="Select Question"
        )
        
        # Add a next question button
        next_button = gr.Button("Next Question")
        
        # Display the question
        gr.Markdown("---")
        question_display = gr.Markdown(value="### Question\n\n" + init_data[6])
        
        with gr.Tabs():
            with gr.TabItem("Game 1: Baseline (A) vs Model (B)"):
                with gr.Row():
                    with gr.Column():
                        gr.Markdown("### Baseline (A)")
                        baseline_answer1 = gr.Markdown(value=init_data[7])
                    with gr.Column():
                        gr.Markdown("### Model (B)")
                        model_answer1 = gr.Markdown(value=init_data[8])
                gr.Markdown("---")
                gr.Markdown("### Judgment")
                judgment1 = gr.Markdown(value=init_data[9])
            
            with gr.TabItem("Game 2: Model (A) vs Baseline (B)"):
                with gr.Row():
                    with gr.Column():
                        gr.Markdown("### Model (A)")
                        model_answer2 = gr.Markdown(value=init_data[8])
                    with gr.Column():
                        gr.Markdown("### Baseline (B)")
                        baseline_answer2 = gr.Markdown(value=init_data[7])
                gr.Markdown("---")
                gr.Markdown("### Judgment")
                judgment2 = gr.Markdown(value=init_data[10])
                
        gr.Markdown("---")
        gr.Markdown("### Citation")
        gr.Markdown("If you find this tool useful, please cite the following papers:")
        gr.Markdown(
            '''```bibtex
@article{li2024crowdsourced,
  title={From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline},
  author={Li, Tianle and Chiang, Wei-Lin and Frick, Evan and Dunlap, Lisa and Wu, Tianhao and Zhu, Banghua and Gonzalez, Joseph E and Stoica, Ion},
  journal={arXiv preprint arXiv:2406.11939},
  year={2024}
}
@misc{arenahard2024,
    title = {From Live Data to High-Quality Benchmarks: The Arena-Hard Pipeline},
    url = {https://lmsys.org/blog/2024-04-19-arena-hard/},
    author = {Tianle Li*, Wei-Lin Chiang*, Evan Frick, Lisa Dunlap, Banghua Zhu, Joseph E. Gonzalez, Ion Stoica},
    month = {April},
    year = {2024}
}
```''')
        
        # Set up event handlers
        benchmark_dropdown.change(
            fn=init_app,
            inputs=benchmark_dropdown,
            outputs=[
                category_dropdown, language_dropdown, judge_dropdown, model_dropdown, 
                question_dropdown, question_dict,
                question_display, 
                baseline_answer1, model_answer1,
                judgment1, judgment2
            ]
        ).then(
            fn=lambda model, baseline: (model, baseline),
            inputs=[model_answer1, baseline_answer1],
            outputs=[model_answer2, baseline_answer2]
        )
        
        # Update questions when category changes
        category_dropdown.change(
            fn=update_questions,
            inputs=[benchmark_dropdown, category_dropdown, language_dropdown],
            outputs=[question_dropdown, question_dict]
        ).then(
            fn=display_content,
            inputs=[benchmark_dropdown, category_dropdown, language_dropdown, judge_dropdown, model_dropdown, question_dropdown, question_dict],
            outputs=[question_display, baseline_answer1, model_answer1, judgment1, judgment2]
        ).then(
            fn=lambda model, baseline: (model, baseline),
            inputs=[model_answer1, baseline_answer1],
            outputs=[model_answer2, baseline_answer2]
        )
        
        # Update questions when language changes
        language_dropdown.change(
            fn=update_questions,
            inputs=[benchmark_dropdown, category_dropdown, language_dropdown],
            outputs=[question_dropdown, question_dict]
        ).then(
            fn=display_content,
            inputs=[benchmark_dropdown, category_dropdown, language_dropdown, judge_dropdown, model_dropdown, question_dropdown, question_dict],
            outputs=[question_display, baseline_answer1, model_answer1, judgment1, judgment2]
        ).then(
            fn=lambda model, baseline: (model, baseline),
            inputs=[model_answer1, baseline_answer1],
            outputs=[model_answer2, baseline_answer2]
        )
        
        # Update models when judge changes
        judge_dropdown.change(
            fn=update_models,
            inputs=[benchmark_dropdown, judge_dropdown],
            outputs=model_dropdown
        ).then(
            fn=display_content,
            inputs=[benchmark_dropdown, category_dropdown, language_dropdown, judge_dropdown, model_dropdown, question_dropdown, question_dict],
            outputs=[question_display, baseline_answer1, model_answer1, judgment1, judgment2]
        ).then(
            fn=lambda model, baseline: (model, baseline),
            inputs=[model_answer1, baseline_answer1],
            outputs=[model_answer2, baseline_answer2]
        )
        
        # Display content when model changes
        model_dropdown.change(
            fn=display_content,
            inputs=[benchmark_dropdown, category_dropdown, language_dropdown, judge_dropdown, model_dropdown, question_dropdown, question_dict],
            outputs=[question_display, baseline_answer1, model_answer1, judgment1, judgment2]
        ).then(
            fn=lambda model, baseline: (model, baseline),
            inputs=[model_answer1, baseline_answer1],
            outputs=[model_answer2, baseline_answer2]
        )
        
        # Display content when question changes
        question_dropdown.change(
            fn=display_content,
            inputs=[benchmark_dropdown, category_dropdown, language_dropdown, judge_dropdown, model_dropdown, question_dropdown, question_dict],
            outputs=[question_display, baseline_answer1, model_answer1, judgment1, judgment2]
        ).then(
            fn=lambda model, baseline: (model, baseline),
            inputs=[model_answer1, baseline_answer1],
            outputs=[model_answer2, baseline_answer2]
        )
        
        # Handle next question button
        next_button.click(
            fn=next_question,
            inputs=[benchmark_dropdown, category_dropdown, language_dropdown, question_dropdown, question_dict],
            outputs=question_dropdown
        )
    
    return app

if __name__ == "__main__":
    import argparse
    
    parser = argparse.ArgumentParser()
    parser.add_argument("--host", type=str, default="0.0.0.0")
    parser.add_argument("--port", type=int)
    parser.add_argument("--share", action="store_true")
    args = parser.parse_args()
    
    app = create_app()
    app.launch(server_name=args.host, server_port=args.port, share=args.share)