File size: 16,663 Bytes
8bedda3
72650c2
c3df5b3
8130dc1
8bedda3
c3df5b3
8bedda3
72650c2
1edf6fb
8bedda3
 
df02c32
 
72650c2
 
 
 
8bedda3
 
1edf6fb
 
 
 
8bedda3
1edf6fb
 
 
 
b97167c
1edf6fb
 
 
72650c2
1edf6fb
 
 
72650c2
1edf6fb
b97167c
1edf6fb
 
 
 
 
 
 
 
7e04c2f
 
1edf6fb
 
 
 
 
 
72650c2
 
8bedda3
72650c2
 
 
 
8bedda3
 
 
 
 
 
72650c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5fb2c04
72650c2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1edf6fb
 
 
 
 
 
 
 
 
 
6ed21d5
1edf6fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c4515ff
1edf6fb
 
 
72650c2
1edf6fb
b97167c
 
1edf6fb
 
 
6ed21d5
1edf6fb
6ed21d5
1edf6fb
 
6ed21d5
1edf6fb
 
 
 
 
 
 
 
 
 
 
 
 
fa37411
 
 
 
 
1edf6fb
 
9b416dc
1edf6fb
72650c2
1edf6fb
72650c2
 
8bedda3
 
 
 
 
72650c2
 
1edf6fb
 
72650c2
1edf6fb
72650c2
 
1edf6fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fa37411
1edf6fb
 
 
 
 
 
 
 
 
fa37411
1edf6fb
 
 
 
8020229
1edf6fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8bedda3
72650c2
8bedda3
1edf6fb
72650c2
1edf6fb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49e21e1
 
1edf6fb
 
 
8bedda3
9b416dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dcb1547
 
 
f0939aa
 
 
 
ed5bd07
 
f0939aa
ed5bd07
 
 
 
b1245be
f0939aa
 
 
 
 
 
 
 
 
ed5bd07
9b416dc
8bedda3
72650c2
 
 
8bedda3
 
72650c2
9b416dc
8bedda3
72650c2
1edf6fb
72650c2
8bedda3
 
 
c3df5b3
 
 
 
 
8130dc1
 
 
 
 
 
 
 
 
5dec4d0
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
"""A gradio app that renders a static leaderboard. This is used for Hugging Face Space."""
import ast
import argparse
import glob
import pickle

import gradio as gr
import numpy as np
import pandas as pd


# notebook_url = "https://colab.research.google.com/drive/1RAWb22-PFNI-X1gPVzc927SGUdfr6nsR?usp=sharing"
notebook_url = "https://colab.research.google.com/drive/1KdwokPjirkTmpO_P1WByFNFiqxWQquwH#scrollTo=o_CpbkGEbhrK"


basic_component_values = [None] * 6
leader_component_values = [None] * 5


def make_default_md(arena_df, elo_results):
    total_votes = sum(arena_df["num_battles"]) // 2
    total_models = len(arena_df)

    leaderboard_md = f"""
# πŸ† LMSYS Chatbot Arena Leaderboard
| [Vote](https://chat.lmsys.org) | [Blog](https://lmsys.org/blog/2023-05-03-arena/) | [GitHub](https://github.com/lm-sys/FastChat) | [Paper](https://arxiv.org/abs/2306.05685) | [Dataset](https://github.com/lm-sys/FastChat/blob/main/docs/dataset_release.md) | [Twitter](https://twitter.com/lmsysorg) | [Discord](https://discord.gg/HSWAKCrnFx) |

LMSYS [Chatbot Arena](https://lmsys.org/blog/2023-05-03-arena/) is a crowdsourced open platform for LLM evals.
We've collected over **400,000** human preference votes to rank LLMs with the Elo ranking system.
"""
    return leaderboard_md


def make_arena_leaderboard_md(arena_df):
    total_votes = sum(arena_df["num_battles"]) // 2
    total_models = len(arena_df)

    leaderboard_md = f"""
Total #models: **{total_models}**. Total #votes: **{total_votes}**. Last updated: March 13, 2024.

Contribute your vote πŸ—³οΈ at [chat.lmsys.org](https://chat.lmsys.org)! Find more analysis in the [notebook]({notebook_url}).
"""
    return leaderboard_md


def make_full_leaderboard_md(elo_results):
    leaderboard_md = f"""
Three benchmarks are displayed: **Arena Elo**, **MT-Bench** and **MMLU**.
- [Chatbot Arena](https://chat.lmsys.org/?arena) - a crowdsourced, randomized battle platform. We use 200K+ user votes to compute Elo ratings.
- [MT-Bench](https://arxiv.org/abs/2306.05685): a set of challenging multi-turn questions. We use GPT-4 to grade the model responses.
- [MMLU](https://arxiv.org/abs/2009.03300) (5-shot): a test to measure a model's multitask accuracy on 57 tasks.

πŸ’» Code: The MT-bench scores (single-answer grading on a scale of 10) are computed by [fastchat.llm_judge](https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge).
The MMLU scores are mostly computed by [InstructEval](https://github.com/declare-lab/instruct-eval).
Higher values are better for all benchmarks. Empty cells mean not available.
"""
    return leaderboard_md


def make_leaderboard_md_live(elo_results):
    leaderboard_md = f"""
# Leaderboard
Last updated: {elo_results["last_updated_datetime"]}
{elo_results["leaderboard_table"]}
"""
    return leaderboard_md


def update_elo_components(max_num_files, elo_results_file):
    log_files = get_log_files(max_num_files)

    # Leaderboard
    if elo_results_file is None:  # Do live update
        battles = clean_battle_data(log_files)
        elo_results = report_elo_analysis_results(battles)

        leader_component_values[0] = make_leaderboard_md_live(elo_results)
        leader_component_values[1] = elo_results["win_fraction_heatmap"]
        leader_component_values[2] = elo_results["battle_count_heatmap"]
        leader_component_values[3] = elo_results["bootstrap_elo_rating"]
        leader_component_values[4] = elo_results["average_win_rate_bar"]

    # Basic stats
    basic_stats = report_basic_stats(log_files)
    md0 = f"Last updated: {basic_stats['last_updated_datetime']}"

    md1 = "### Action Histogram\n"
    md1 += basic_stats["action_hist_md"] + "\n"

    md2 = "### Anony. Vote Histogram\n"
    md2 += basic_stats["anony_vote_hist_md"] + "\n"

    md3 = "### Model Call Histogram\n"
    md3 += basic_stats["model_hist_md"] + "\n"

    md4 = "### Model Call (Last 24 Hours)\n"
    md4 += basic_stats["num_chats_last_24_hours"] + "\n"

    basic_component_values[0] = md0
    basic_component_values[1] = basic_stats["chat_dates_bar"]
    basic_component_values[2] = md1
    basic_component_values[3] = md2
    basic_component_values[4] = md3
    basic_component_values[5] = md4


def update_worker(max_num_files, interval, elo_results_file):
    while True:
        tic = time.time()
        update_elo_components(max_num_files, elo_results_file)
        durtaion = time.time() - tic
        print(f"update duration: {durtaion:.2f} s")
        time.sleep(max(interval - durtaion, 0))


def load_demo(url_params, request: gr.Request):
    logger.info(f"load_demo. ip: {request.client.host}. params: {url_params}")
    return basic_component_values + leader_component_values


def model_hyperlink(model_name, link):
    return f'<a target="_blank" href="{link}" style="color: var(--link-text-color); text-decoration: underline;text-decoration-style: dotted;">{model_name}</a>'


def load_leaderboard_table_csv(filename, add_hyperlink=True):
    lines = open(filename).readlines()
    heads = [v.strip() for v in lines[0].split(",")]
    rows = []
    for i in range(1, len(lines)):
        row = [v.strip() for v in lines[i].split(",")]
        for j in range(len(heads)):
            item = {}
            for h, v in zip(heads, row):
                if h == "Arena Elo rating":
                    if v != "-":
                        v = int(ast.literal_eval(v))
                    else:
                        v = np.nan
                elif h == "MMLU":
                    if v != "-":
                        v = round(ast.literal_eval(v) * 100, 1)
                    else:
                        v = np.nan
                elif h == "MT-bench (win rate %)":
                    if v != "-":
                        v = round(ast.literal_eval(v[:-1]), 1)
                    else:
                        v = np.nan
                elif h == "MT-bench (score)":
                    if v != "-":
                        v = round(ast.literal_eval(v), 2)
                    else:
                        v = np.nan
                item[h] = v
            if add_hyperlink:
                item["Model"] = model_hyperlink(item["Model"], item["Link"])
        rows.append(item)

    return rows


def build_basic_stats_tab():
    empty = "Loading ..."
    basic_component_values[:] = [empty, None, empty, empty, empty, empty]

    md0 = gr.Markdown(empty)
    gr.Markdown("#### Figure 1: Number of model calls and votes")
    plot_1 = gr.Plot(show_label=False)
    with gr.Row():
        with gr.Column():
            md1 = gr.Markdown(empty)
        with gr.Column():
            md2 = gr.Markdown(empty)
    with gr.Row():
        with gr.Column():
            md3 = gr.Markdown(empty)
        with gr.Column():
            md4 = gr.Markdown(empty)
    return [md0, plot_1, md1, md2, md3, md4]

def get_full_table(arena_df, model_table_df):
    values = []
    for i in range(len(model_table_df)):
        row = []
        model_key = model_table_df.iloc[i]["key"]
        model_name = model_table_df.iloc[i]["Model"]
        # model display name
        row.append(model_name)
        if model_key in arena_df.index:
            idx = arena_df.index.get_loc(model_key)
            row.append(round(arena_df.iloc[idx]["rating"]))
        else:
            row.append(np.nan)
        row.append(model_table_df.iloc[i]["MT-bench (score)"])
        row.append(model_table_df.iloc[i]["MMLU"])
        # Organization
        row.append(model_table_df.iloc[i]["Organization"])
        # license
        row.append(model_table_df.iloc[i]["License"])

        values.append(row)
    values.sort(key=lambda x: -x[1] if not np.isnan(x[1]) else 1e9)
    return values


def get_arena_table(arena_df, model_table_df):
    # sort by rating
    arena_df = arena_df.sort_values(by=["rating"], ascending=False)
    values = []
    for i in range(len(arena_df)):
        row = []
        model_key = arena_df.index[i]
        print(model_key)
        model_name = model_table_df[model_table_df["key"] == model_key]["Model"].values[
            0
        ]

        # rank
        ranking = arena_df.iloc[i].get("final_ranking") or i+1
        row.append(ranking)
        # model display name
        row.append(model_name)
        # elo rating
        row.append(round(arena_df.iloc[i]["rating"]))
        upper_diff = round(
            arena_df.iloc[i]["rating_q975"] - arena_df.iloc[i]["rating"]
        )
        lower_diff = round(
            arena_df.iloc[i]["rating"] - arena_df.iloc[i]["rating_q025"]
        )
        row.append(f"+{upper_diff}/-{lower_diff}")
        # num battles
        row.append(round(arena_df.iloc[i]["num_battles"]))
        # Organization
        row.append(
            model_table_df[model_table_df["key"] == model_key]["Organization"].values[0]
        )
        # license
        row.append(
            model_table_df[model_table_df["key"] == model_key]["License"].values[0]
        )

        cutoff_date = model_table_df[model_table_df["key"] == model_key]["Knowledge cutoff date"].values[0]
        if cutoff_date == "-":
            row.append("Unknown")
        else:
            row.append(cutoff_date)
        values.append(row)
    return values

def build_leaderboard_tab(elo_results_file, leaderboard_table_file, show_plot=False):
    if elo_results_file is None:  # Do live update
        default_md = "Loading ..."
        p1 = p2 = p3 = p4 = None
    else:
        with open(elo_results_file, "rb") as fin:
            elo_results = pickle.load(fin)

        p1 = elo_results["win_fraction_heatmap"]
        p2 = elo_results["battle_count_heatmap"]
        p3 = elo_results["bootstrap_elo_rating"]
        p4 = elo_results["average_win_rate_bar"]
        arena_df = elo_results["leaderboard_table_df"]
        default_md = make_default_md(arena_df, elo_results)

    md_1 = gr.Markdown(default_md, elem_id="leaderboard_markdown")
    if leaderboard_table_file:
        data = load_leaderboard_table_csv(leaderboard_table_file)
        model_table_df = pd.DataFrame(data)

        with gr.Tabs() as tabs:
            # arena table
            arena_table_vals = get_arena_table(arena_df, model_table_df)
            with gr.Tab("Arena Elo", id=0):
                md = make_arena_leaderboard_md(arena_df)
                gr.Markdown(md, elem_id="leaderboard_markdown")
                gr.Dataframe(
                    headers=[
                        "Rank",
                        "πŸ€– Model",
                        "⭐ Arena Elo",
                        "πŸ“Š 95% CI",
                        "πŸ—³οΈ Votes",
                        "Organization",
                        "License",
                        "Knowledge Cutoff",
                    ],
                    datatype=[
                        "str",
                        "markdown",
                        "number",
                        "str",
                        "number",
                        "str",
                        "str",
                        "str",
                    ],
                    value=arena_table_vals,
                    elem_id="arena_leaderboard_dataframe",
                    height=700,
                    column_widths=[50, 200, 120, 100, 100, 150, 150, 100],
                    wrap=True,
                )
            with gr.Tab("Full Leaderboard", id=1):
                md = make_full_leaderboard_md(elo_results)
                gr.Markdown(md, elem_id="leaderboard_markdown")
                full_table_vals = get_full_table(arena_df, model_table_df)
                gr.Dataframe(
                    headers=[
                        "πŸ€– Model",
                        "⭐ Arena Elo",
                        "πŸ“ˆ MT-bench",
                        "πŸ“š MMLU",
                        "Organization",
                        "License",
                    ],
                    datatype=["markdown", "number", "number", "number", "str", "str"],
                    value=full_table_vals,
                    elem_id="full_leaderboard_dataframe",
                    column_widths=[200, 100, 100, 100, 150, 150],
                    height=700,
                    wrap=True,
                )
        if not show_plot:
            gr.Markdown(
                """ ## Visit our [HF space](https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard) for more analysis!
                If you want to see more models, please help us [add them](https://github.com/lm-sys/FastChat/blob/main/docs/arena.md#how-to-add-a-new-model).
                """,
                elem_id="leaderboard_markdown",
            )
    else:
        pass

    leader_component_values[:] = [default_md, p1, p2, p3, p4]

    if show_plot:
        gr.Markdown(
            f"""## More Statistics for Chatbot Arena\n
Below are figures for more statistics. The code for generating them is also included in this [notebook]({notebook_url}).
You can find more discussions in this blog [post](https://lmsys.org/blog/2023-12-07-leaderboard/).
    """,
            elem_id="leaderboard_markdown"
        )
        with gr.Row():
            with gr.Column():
                gr.Markdown(
                    "#### Figure 1: Fraction of Model A Wins for All Non-tied A vs. B Battles"
                )
                plot_1 = gr.Plot(p1, show_label=False)
            with gr.Column():
                gr.Markdown(
                    "#### Figure 2: Battle Count for Each Combination of Models (without Ties)"
                )
                plot_2 = gr.Plot(p2, show_label=False)
        with gr.Row():
            with gr.Column():
                gr.Markdown(
                    "#### Figure 3: Bootstrap of Elo Estimates (1000 Rounds of Random Sampling)"
                )
                plot_3 = gr.Plot(p3, show_label=False)
            with gr.Column():
                gr.Markdown(
                    "#### Figure 4: Average Win Rate Against All Other Models (Assuming Uniform Sampling and No Ties)"
                )
                plot_4 = gr.Plot(p4, show_label=False)

    gr.Markdown(acknowledgment_md)

    if show_plot:
        return [md_1, plot_1, plot_2, plot_3, plot_4]
    return [md_1]

block_css = """
#notice_markdown {
    font-size: 104%
}
#notice_markdown th {
    display: none;
}
#notice_markdown td {
    padding-top: 6px;
    padding-bottom: 6px;
}
#leaderboard_markdown {
    font-size: 104%
}
#leaderboard_markdown td {
    padding-top: 6px;
    padding-bottom: 6px;
}
#leaderboard_dataframe td {
    line-height: 0.1em;
}
footer {
    display:none !important
}
.sponsor-image-about img {
    margin: 0 20px;
    margin-top: 20px;
    height: 40px;
    max-height: 100%;
    width: auto;
    float: left;
}
"""

acknowledgment_md = """
### Acknowledgment
We thank [Kaggle](https://www.kaggle.com/), [MBZUAI](https://mbzuai.ac.ae/), [a16z](https://www.a16z.com/), [Together AI](https://www.together.ai/), [Anyscale](https://www.anyscale.com/), [HuggingFace](https://huggingface.co/) for their generous [sponsorship](https://lmsys.org/donations/).

<div class="sponsor-image-about">
    <img src="https://storage.googleapis.com/public-arena-asset/kaggle.png" alt="Kaggle">
    <img src="https://storage.googleapis.com/public-arena-asset/mbzuai.jpeg" alt="MBZUAI">
    <img src="https://storage.googleapis.com/public-arena-asset/a16z.jpeg" alt="a16z">
    <img src="https://storage.googleapis.com/public-arena-asset/together.png" alt="Together AI">
    <img src="https://storage.googleapis.com/public-arena-asset/anyscale.png" alt="AnyScale">
    <img src="https://storage.googleapis.com/public-arena-asset/huggingface.png" alt="HuggingFace">
</div>
"""

def build_demo(elo_results_file, leaderboard_table_file):
    text_size = gr.themes.sizes.text_lg

    with gr.Blocks(
        title="Chatbot Arena Leaderboard",
        theme=gr.themes.Base(text_size=text_size),
        css=block_css,
    ) as demo:
        leader_components = build_leaderboard_tab(
            elo_results_file, leaderboard_table_file, show_plot=True
        )
    return demo


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--share", action="store_true")
    args = parser.parse_args()

    elo_result_files = glob.glob("elo_results_*.pkl")
    elo_result_files.sort(key=lambda x: int(x[12:-4]))
    elo_result_file = elo_result_files[-1]

    leaderboard_table_files = glob.glob("leaderboard_table_*.csv")
    leaderboard_table_files.sort(key=lambda x: int(x[18:-4]))
    leaderboard_table_file = leaderboard_table_files[-1]

    demo = build_demo(elo_result_file, leaderboard_table_file)
    demo.launch(share=args.share)