File size: 18,838 Bytes
f3305db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
"""
Chatbot Arena (battle) tab.
Users chat with two anonymous models.
"""

import json
import time

import gradio as gr
import numpy as np

from fastchat.constants import (
    MODERATION_MSG,
    CONVERSATION_LIMIT_MSG,
    SLOW_MODEL_MSG,
    BLIND_MODE_INPUT_CHAR_LEN_LIMIT,
    CONVERSATION_TURN_LIMIT,
    SURVEY_LINK,
)
from fastchat.model.model_adapter import get_conversation_template
from fastchat.serve.gradio_block_arena_named import flash_buttons
from fastchat.serve.gradio_web_server import (
    State,
    bot_response,
    get_conv_log_filename,
    no_change_btn,
    enable_btn,
    disable_btn,
    invisible_btn,
    enable_text,
    disable_text,
    acknowledgment_md,
    get_ip,
    get_model_description_md,
)
from fastchat.serve.remote_logger import get_remote_logger
from fastchat.utils import (
    build_logger,
    moderation_filter,
)

logger = build_logger("gradio_web_server_multi", "gradio_web_server_multi.log")

num_sides = 2
enable_moderation = False
anony_names = ["", ""]
models = []


def set_global_vars_anony(enable_moderation_):
    global enable_moderation
    enable_moderation = enable_moderation_


def load_demo_side_by_side_anony(models_, url_params):
    global models
    models = models_

    states = (None,) * num_sides
    selector_updates = (
        gr.Markdown(visible=True),
        gr.Markdown(visible=True),
    )

    return states + selector_updates


def vote_last_response(states, vote_type, model_selectors, request: gr.Request):
    with open(get_conv_log_filename(), "a") as fout:
        data = {
            "tstamp": round(time.time(), 4),
            "type": vote_type,
            "models": [x for x in model_selectors],
            "states": [x.dict() for x in states],
            "ip": get_ip(request),
        }
        fout.write(json.dumps(data) + "\n")
    get_remote_logger().log(data)

    gr.Info(
        "πŸŽ‰ Thanks for voting! Your vote shapes the leaderboard, please vote RESPONSIBLY."
    )
    if ":" not in model_selectors[0]:
        for i in range(5):
            names = (
                "### Model A: " + states[0].model_name,
                "### Model B: " + states[1].model_name,
            )
            # yield names + ("",) + (disable_btn,) * 4
            yield names + (disable_text,) + (disable_btn,) * 5
            time.sleep(0.1)
    else:
        names = (
            "### Model A: " + states[0].model_name,
            "### Model B: " + states[1].model_name,
        )
        # yield names + ("",) + (disable_btn,) * 4
        yield names + (disable_text,) + (disable_btn,) * 5


def leftvote_last_response(
    state0, state1, model_selector0, model_selector1, request: gr.Request
):
    logger.info(f"leftvote (anony). ip: {get_ip(request)}")
    for x in vote_last_response(
        [state0, state1], "leftvote", [model_selector0, model_selector1], request
    ):
        yield x


def rightvote_last_response(
    state0, state1, model_selector0, model_selector1, request: gr.Request
):
    logger.info(f"rightvote (anony). ip: {get_ip(request)}")
    for x in vote_last_response(
        [state0, state1], "rightvote", [model_selector0, model_selector1], request
    ):
        yield x


def tievote_last_response(
    state0, state1, model_selector0, model_selector1, request: gr.Request
):
    logger.info(f"tievote (anony). ip: {get_ip(request)}")
    for x in vote_last_response(
        [state0, state1], "tievote", [model_selector0, model_selector1], request
    ):
        yield x


def bothbad_vote_last_response(
    state0, state1, model_selector0, model_selector1, request: gr.Request
):
    logger.info(f"bothbad_vote (anony). ip: {get_ip(request)}")
    for x in vote_last_response(
        [state0, state1], "bothbad_vote", [model_selector0, model_selector1], request
    ):
        yield x


def regenerate(state0, state1, request: gr.Request):
    logger.info(f"regenerate (anony). ip: {get_ip(request)}")
    states = [state0, state1]
    if state0.regen_support and state1.regen_support:
        for i in range(num_sides):
            states[i].conv.update_last_message(None)
        return (
            states + [x.to_gradio_chatbot() for x in states] + [""] + [disable_btn] * 6
        )
    states[0].skip_next = True
    states[1].skip_next = True
    return states + [x.to_gradio_chatbot() for x in states] + [""] + [no_change_btn] * 6


def clear_history(request: gr.Request):
    logger.info(f"clear_history (anony). ip: {get_ip(request)}")
    return (
        [None] * num_sides
        + [None] * num_sides
        + anony_names
        + [enable_text]
        + [invisible_btn] * 4
        + [disable_btn] * 2
        + [""]
        + [enable_btn]
    )


def share_click(state0, state1, model_selector0, model_selector1, request: gr.Request):
    logger.info(f"share (anony). ip: {get_ip(request)}")
    if state0 is not None and state1 is not None:
        vote_last_response(
            [state0, state1], "share", [model_selector0, model_selector1], request
        )


SAMPLING_WEIGHTS = {}

# target model sampling weights will be boosted.
BATTLE_TARGETS = {}

ANON_MODELS = []

SAMPLING_BOOST_MODELS = []

# outage models won't be sampled.
OUTAGE_MODELS = []


def get_sample_weight(model, outage_models, sampling_weights, sampling_boost_models=[]):
    if model in outage_models:
        return 0
    weight = sampling_weights.get(model, 0)
    if model in sampling_boost_models:
        weight *= 5
    return weight


def get_battle_pair(
    models, battle_targets, outage_models, sampling_weights, sampling_boost_models
):
    if len(models) == 1:
        return models[0], models[0]

    model_weights = []
    for model in models:
        weight = get_sample_weight(
            model, outage_models, sampling_weights, sampling_boost_models
        )
        model_weights.append(weight)
    total_weight = np.sum(model_weights)
    model_weights = model_weights / total_weight
    chosen_idx = np.random.choice(len(models), p=model_weights)
    chosen_model = models[chosen_idx]
    # for p, w in zip(models, model_weights):
    #     print(p, w)

    rival_models = []
    rival_weights = []
    for model in models:
        if model == chosen_model:
            continue
        if model in ANON_MODELS and chosen_model in ANON_MODELS:
            continue
        weight = get_sample_weight(model, outage_models, sampling_weights)
        if (
            weight != 0
            and chosen_model in battle_targets
            and model in battle_targets[chosen_model]
        ):
            # boost to 20% chance
            weight = 0.5 * total_weight / len(battle_targets[chosen_model])
        rival_models.append(model)
        rival_weights.append(weight)
    # for p, w in zip(rival_models, rival_weights):
    #     print(p, w)
    rival_weights = rival_weights / np.sum(rival_weights)
    rival_idx = np.random.choice(len(rival_models), p=rival_weights)
    rival_model = rival_models[rival_idx]

    swap = np.random.randint(2)
    if swap == 0:
        return chosen_model, rival_model
    else:
        return rival_model, chosen_model


def add_text(
    state0, state1, model_selector0, model_selector1, text, request: gr.Request
):
    ip = get_ip(request)
    logger.info(f"add_text (anony). ip: {ip}. len: {len(text)}")
    states = [state0, state1]
    model_selectors = [model_selector0, model_selector1]

    # Init states if necessary
    if states[0] is None:
        assert states[1] is None

        model_left, model_right = get_battle_pair(
            models,
            BATTLE_TARGETS,
            OUTAGE_MODELS,
            SAMPLING_WEIGHTS,
            SAMPLING_BOOST_MODELS,
        )
        states = [
            State(model_left),
            State(model_right),
        ]

    if len(text) <= 0:
        for i in range(num_sides):
            states[i].skip_next = True
        return (
            states
            + [x.to_gradio_chatbot() for x in states]
            + ["", None]
            + [
                no_change_btn,
            ]
            * 6
            + [""]
        )

    model_list = [states[i].model_name for i in range(num_sides)]
    # turn on moderation in battle mode
    all_conv_text_left = states[0].conv.get_prompt()
    all_conv_text_right = states[0].conv.get_prompt()
    all_conv_text = (
        all_conv_text_left[-1000:] + all_conv_text_right[-1000:] + "\nuser: " + text
    )
    flagged = moderation_filter(all_conv_text, model_list, do_moderation=True)
    if flagged:
        logger.info(f"violate moderation (anony). ip: {ip}. text: {text}")
        # overwrite the original text
        text = MODERATION_MSG

    conv = states[0].conv
    if (len(conv.messages) - conv.offset) // 2 >= CONVERSATION_TURN_LIMIT:
        logger.info(f"conversation turn limit. ip: {get_ip(request)}. text: {text}")
        for i in range(num_sides):
            states[i].skip_next = True
        return (
            states
            + [x.to_gradio_chatbot() for x in states]
            + [CONVERSATION_LIMIT_MSG]
            + [
                no_change_btn,
            ]
            * 6
            + [""]
        )

    text = text[:BLIND_MODE_INPUT_CHAR_LEN_LIMIT]  # Hard cut-off
    for i in range(num_sides):
        states[i].conv.append_message(states[i].conv.roles[0], text)
        states[i].conv.append_message(states[i].conv.roles[1], None)
        states[i].skip_next = False

    hint_msg = ""
    for i in range(num_sides):
        if "deluxe" in states[i].model_name:
            hint_msg = SLOW_MODEL_MSG
    return (
        states
        + [x.to_gradio_chatbot() for x in states]
        + [""]
        + [
            disable_btn,
        ]
        * 6
        + [hint_msg]
    )


def bot_response_multi(
    state0,
    state1,
    temperature,
    top_p,
    max_new_tokens,
    request: gr.Request,
):
    logger.info(f"bot_response_multi (anony). ip: {get_ip(request)}")

    if state0 is None or state0.skip_next:
        # This generate call is skipped due to invalid inputs
        yield (
            state0,
            state1,
            state0.to_gradio_chatbot(),
            state1.to_gradio_chatbot(),
        ) + (no_change_btn,) * 6
        return

    states = [state0, state1]
    gen = []
    for i in range(num_sides):
        gen.append(
            bot_response(
                states[i],
                temperature,
                top_p,
                max_new_tokens,
                request,
                apply_rate_limit=False,
                use_recommended_config=True,
            )
        )

    model_tpy = []
    for i in range(num_sides):
        token_per_yield = 1
        if states[i].model_name in [
            "gemini-pro",
            "gemma-1.1-2b-it",
            "gemma-1.1-7b-it",
            "phi-3-mini-4k-instruct",
            "phi-3-mini-128k-instruct",
            "snowflake-arctic-instruct",
        ]:
            token_per_yield = 30
        elif states[i].model_name in [
            "qwen-max-0428",
            "qwen1.5-110b-chat",
            "llava-v1.6-34b",
        ]:
            token_per_yield = 7
        elif states[i].model_name in [
            "qwen2-72b-instruct",
        ]:
            token_per_yield = 4
        model_tpy.append(token_per_yield)

    chatbots = [None] * num_sides
    iters = 0
    while True:
        stop = True
        iters += 1
        for i in range(num_sides):
            try:
                # yield fewer times if chunk size is larger
                if model_tpy[i] == 1 or (iters % model_tpy[i] == 1 or iters < 3):
                    ret = next(gen[i])
                    states[i], chatbots[i] = ret[0], ret[1]
                stop = False
            except StopIteration:
                pass
        yield states + chatbots + [disable_btn] * 6
        if stop:
            break


def build_side_by_side_ui_anony(models):
    notice_markdown = f"""
# βš”οΈ  LMSYS Chatbot Arena: Benchmarking LLMs in the Wild
[Blog](https://lmsys.org/blog/2023-05-03-arena/) | [GitHub](https://github.com/lm-sys/FastChat) | [Paper](https://arxiv.org/abs/2403.04132) | [Dataset](https://github.com/lm-sys/FastChat/blob/main/docs/dataset_release.md) | [Twitter](https://twitter.com/lmsysorg) | [Discord](https://discord.gg/HSWAKCrnFx) | [Kaggle Competition](https://www.kaggle.com/competitions/lmsys-chatbot-arena)

{SURVEY_LINK}

## πŸ“£ News
- Chatbot Arena now supports images in beta. Check it out [here](https://chat.lmsys.org/?vision).

## πŸ“œ Rules
- Ask any question to two anonymous models (e.g., ChatGPT, Gemini, Claude, Llama) and vote for the better one!
- You can chat for multiple turns until you identify a winner.
- Votes won't be counted if model identities are revealed during the conversation.

## πŸ† Chatbot Arena [Leaderboard](https://leaderboard.lmsys.org)
- We've collected **1,000,000+** human votes to compute an LLM leaderboard for 100+ models. Find out who is the πŸ₯‡LLM Champion [here](https://leaderboard.lmsys.org)!

## πŸ‘‡ Chat now!
"""

    states = [gr.State() for _ in range(num_sides)]
    model_selectors = [None] * num_sides
    chatbots = [None] * num_sides

    gr.Markdown(notice_markdown, elem_id="notice_markdown")

    with gr.Group(elem_id="share-region-anony"):
        with gr.Accordion(
            f"πŸ” Expand to see the descriptions of {len(models)} models", open=False
        ):
            model_description_md = get_model_description_md(models)
            gr.Markdown(model_description_md, elem_id="model_description_markdown")
        with gr.Row():
            for i in range(num_sides):
                label = "Model A" if i == 0 else "Model B"
                with gr.Column():
                    chatbots[i] = gr.Chatbot(
                        label=label,
                        elem_id="chatbot",
                        height=650,
                        show_copy_button=True,
                    )

        with gr.Row():
            for i in range(num_sides):
                with gr.Column():
                    model_selectors[i] = gr.Markdown(
                        anony_names[i], elem_id="model_selector_md"
                    )
        with gr.Row():
            slow_warning = gr.Markdown("")

    with gr.Row():
        leftvote_btn = gr.Button(
            value="πŸ‘ˆ  A is better", visible=False, interactive=False
        )
        rightvote_btn = gr.Button(
            value="πŸ‘‰  B is better", visible=False, interactive=False
        )
        tie_btn = gr.Button(value="🀝  Tie", visible=False, interactive=False)
        bothbad_btn = gr.Button(
            value="πŸ‘Ž  Both are bad", visible=False, interactive=False
        )

    with gr.Row():
        textbox = gr.Textbox(
            show_label=False,
            placeholder="πŸ‘‰ Enter your prompt and press ENTER",
            elem_id="input_box",
        )
        send_btn = gr.Button(value="Send", variant="primary", scale=0)

    with gr.Row() as button_row:
        clear_btn = gr.Button(value="🎲 New Round", interactive=False)
        regenerate_btn = gr.Button(value="πŸ”„  Regenerate", interactive=False)
        share_btn = gr.Button(value="πŸ“·  Share")

    with gr.Accordion("Parameters", open=False, visible=False) as parameter_row:
        temperature = gr.Slider(
            minimum=0.0,
            maximum=1.0,
            value=0.7,
            step=0.1,
            interactive=True,
            label="Temperature",
        )
        top_p = gr.Slider(
            minimum=0.0,
            maximum=1.0,
            value=1.0,
            step=0.1,
            interactive=True,
            label="Top P",
        )
        max_output_tokens = gr.Slider(
            minimum=16,
            maximum=2048,
            value=1600,
            step=64,
            interactive=True,
            label="Max output tokens",
        )

    gr.Markdown(acknowledgment_md, elem_id="ack_markdown")

    # Register listeners
    btn_list = [
        leftvote_btn,
        rightvote_btn,
        tie_btn,
        bothbad_btn,
        regenerate_btn,
        clear_btn,
    ]
    leftvote_btn.click(
        leftvote_last_response,
        states + model_selectors,
        model_selectors
        + [textbox, leftvote_btn, rightvote_btn, tie_btn, bothbad_btn, send_btn],
    )
    rightvote_btn.click(
        rightvote_last_response,
        states + model_selectors,
        model_selectors
        + [textbox, leftvote_btn, rightvote_btn, tie_btn, bothbad_btn, send_btn],
    )
    tie_btn.click(
        tievote_last_response,
        states + model_selectors,
        model_selectors
        + [textbox, leftvote_btn, rightvote_btn, tie_btn, bothbad_btn, send_btn],
    )
    bothbad_btn.click(
        bothbad_vote_last_response,
        states + model_selectors,
        model_selectors
        + [textbox, leftvote_btn, rightvote_btn, tie_btn, bothbad_btn, send_btn],
    )
    regenerate_btn.click(
        regenerate, states, states + chatbots + [textbox] + btn_list
    ).then(
        bot_response_multi,
        states + [temperature, top_p, max_output_tokens],
        states + chatbots + btn_list,
    ).then(
        flash_buttons, [], btn_list
    )
    clear_btn.click(
        clear_history,
        None,
        states
        + chatbots
        + model_selectors
        + [textbox]
        + btn_list
        + [slow_warning]
        + [send_btn],
    )

    share_js = """
function (a, b, c, d) {
    const captureElement = document.querySelector('#share-region-anony');
    html2canvas(captureElement)
        .then(canvas => {
            canvas.style.display = 'none'
            document.body.appendChild(canvas)
            return canvas
        })
        .then(canvas => {
            const image = canvas.toDataURL('image/png')
            const a = document.createElement('a')
            a.setAttribute('download', 'chatbot-arena.png')
            a.setAttribute('href', image)
            a.click()
            canvas.remove()
        });
    return [a, b, c, d];
}
"""
    share_btn.click(share_click, states + model_selectors, [], js=share_js)

    textbox.submit(
        add_text,
        states + model_selectors + [textbox],
        states + chatbots + [textbox] + btn_list + [slow_warning],
    ).then(
        bot_response_multi,
        states + [temperature, top_p, max_output_tokens],
        states + chatbots + btn_list,
    ).then(
        flash_buttons,
        [],
        btn_list,
    )

    send_btn.click(
        add_text,
        states + model_selectors + [textbox],
        states + chatbots + [textbox] + btn_list,
    ).then(
        bot_response_multi,
        states + [temperature, top_p, max_output_tokens],
        states + chatbots + btn_list,
    ).then(
        flash_buttons, [], btn_list
    )

    return states + model_selectors