File size: 8,525 Bytes
a19f11e
 
 
 
 
 
 
 
076f69b
 
 
65566f3
a19f11e
3b5ccf8
a19f11e
 
50a5912
076f69b
d8b3fa4
 
 
076f69b
65566f3
3b5ccf8
65566f3
 
740b104
 
a19f11e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3b5ccf8
 
 
 
076f69b
3b5ccf8
 
 
 
 
 
 
 
 
a19f11e
076f69b
65566f3
 
740b104
65566f3
 
 
740b104
65566f3
 
a19f11e
65566f3
a19f11e
65566f3
3b5ccf8
 
 
 
 
a19f11e
 
 
 
 
 
 
 
 
 
519572d
 
 
a19f11e
 
 
 
8f146f1
 
 
 
 
 
 
 
 
 
 
 
 
a19f11e
 
65566f3
 
076f69b
65566f3
740b104
 
 
65566f3
076f69b
 
3b5ccf8
 
 
65566f3
076f69b
 
 
a19f11e
076f69b
740b104
3b5ccf8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a19f11e
 
65566f3
b272a27
 
076f69b
a19f11e
076f69b
740b104
65566f3
 
740b104
65566f3
 
 
 
740b104
65566f3
 
 
 
3b5ccf8
 
 
 
 
 
 
 
 
 
65566f3
3b5ccf8
 
 
 
 
 
65566f3
3b5ccf8
 
 
 
 
 
65566f3
740b104
 
 
 
 
 
 
 
 
a19f11e
 
65566f3
b272a27
 
076f69b
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
"""
It provides a leaderboard component.
"""

from collections import defaultdict
import enum
import math

import firebase_admin
from firebase_admin import credentials
from firebase_admin import firestore
from google.cloud.firestore_v1 import base_query
import gradio as gr
import lingua
import pandas as pd

from credentials import get_credentials_json

if gr.NO_RELOAD:
  firebase_admin.initialize_app(credentials.Certificate(get_credentials_json()))
  db = firestore.client()

SUPPORTED_TRANSLATION_LANGUAGES = [
    language.name.capitalize() for language in lingua.Language.all()
]

ANY_LANGUAGE = "Any"


class LeaderboardTab(enum.Enum):
  SUMMARIZATION = "Summarization"
  TRANSLATION = "Translation"


# Ref: https://colab.research.google.com/drive/1RAWb22-PFNI-X1gPVzc927SGUdfr6nsR?usp=sharing#scrollTo=QLGc6DwxyvQc pylint: disable=line-too-long
def compute_elo(battles, k=4, scale=400, base=10, initial_rating=1000):
  rating = defaultdict(lambda: initial_rating)

  for model_a, model_b, winner in battles[["model_a", "model_b",
                                           "winner"]].itertuples(index=False):
    rating_a = rating[model_a]
    rating_b = rating[model_b]

    expected_score_a = 1 / (1 + base**((rating_b - rating_a) / scale))
    expected_score_b = 1 / (1 + base**((rating_a - rating_b) / scale))

    scored_point_a = 0.5 if winner == "tie" else int(winner == "model_a")

    rating[model_a] += k * (scored_point_a - expected_score_a)
    rating[model_b] += k * (1 - scored_point_a - expected_score_b)

  return rating


def get_docs(tab: str,
             summary_lang: str = None,
             source_lang: str = None,
             target_lang: str = None):
  if tab == LeaderboardTab.SUMMARIZATION:
    collection = db.collection("arena-summarizations").order_by("timestamp")

    if summary_lang:
      collection = collection.where(filter=base_query.FieldFilter(
          "model_a_response_language", "==", summary_lang.lower())).where(
              filter=base_query.FieldFilter("model_b_response_language", "==",
                                            summary_lang.lower()))

    return collection.stream()

  if tab == LeaderboardTab.TRANSLATION:
    collection = db.collection("arena-translations").order_by("timestamp")

    if source_lang and (not source_lang == ANY_LANGUAGE):
      collection = collection.where(filter=base_query.FieldFilter(
          "source_language", "==", source_lang.lower()))

    if target_lang and (not target_lang == ANY_LANGUAGE):
      collection = collection.where(filter=base_query.FieldFilter(
          "target_language", "==", target_lang.lower()))

    return collection.stream()


def load_elo_ratings(tab,
                     summary_lang: str = None,
                     source_lang: str = None,
                     target_lang: str = None):
  docs = get_docs(tab, summary_lang, source_lang, target_lang)

  battles = []
  for doc in docs:
    data = doc.to_dict()
    battles.append({
        "model_a": data["model_a"],
        "model_b": data["model_b"],
        "winner": data["winner"]
    })

  if not battles:
    return

  battles = pd.DataFrame(battles)
  ratings = compute_elo(battles)

  sorted_ratings = sorted(ratings.items(), key=lambda x: x[1], reverse=True)

  rank = 0
  last_rating = None
  rating_rows = []
  for index, (model, rating) in enumerate(sorted_ratings):
    int_rating = math.floor(rating + 0.5)
    if int_rating != last_rating:
      rank = index + 1

    rating_rows.append([rank, model, int_rating])
    last_rating = int_rating

  return rating_rows


LEADERBOARD_UPDATE_INTERVAL = 600  # 10 minutes
LEADERBOARD_INFO = "The leaderboard is updated every 10 minutes."

DEFAULT_FILTER_OPTIONS = {
    "summary_language": lingua.Language.ENGLISH.name.capitalize(),
    "source_language": ANY_LANGUAGE,
    "target_language": lingua.Language.ENGLISH.name.capitalize()
}


def update_filtered_leaderboard(tab, summary_lang: str, source_lang: str,
                                target_lang: str):
  new_value = load_elo_ratings(tab, summary_lang, source_lang, target_lang)
  return gr.update(value=new_value)


def build_leaderboard():
  with gr.Tabs():
    with gr.Tab(LeaderboardTab.SUMMARIZATION.value):
      with gr.Accordion("Filter", open=False) as summarization_filter:
        with gr.Row():
          languages = [
              language.name.capitalize() for language in lingua.Language.all()
          ]
          summary_language = gr.Dropdown(
              choices=languages,
              value=DEFAULT_FILTER_OPTIONS["summary_language"],
              label="Summary language",
              interactive=True)

        with gr.Row():
          filtered_summarization = gr.DataFrame(
              headers=["Rank", "Model", "Elo rating"],
              datatype=["number", "str", "number"],
              value=lambda: load_elo_ratings(
                  LeaderboardTab.SUMMARIZATION, DEFAULT_FILTER_OPTIONS[
                      "summary_language"]),
              elem_classes="leaderboard")

      summary_language.change(fn=update_filtered_leaderboard,
                              inputs=[
                                  gr.State(LeaderboardTab.SUMMARIZATION),
                                  summary_language,
                                  gr.State(),
                                  gr.State()
                              ],
                              outputs=filtered_summarization)

      gr.Dataframe(headers=["Rank", "Model", "Elo rating"],
                   datatype=["number", "str", "number"],
                   value=lambda: load_elo_ratings(LeaderboardTab.SUMMARIZATION),
                   every=LEADERBOARD_UPDATE_INTERVAL,
                   elem_classes="leaderboard")
      gr.Markdown(LEADERBOARD_INFO)

    with gr.Tab(LeaderboardTab.TRANSLATION.value):
      with gr.Accordion("Filter", open=False) as translation_filter:
        with gr.Row():
          source_language = gr.Dropdown(
              choices=SUPPORTED_TRANSLATION_LANGUAGES + [ANY_LANGUAGE],
              label="Source language",
              value=DEFAULT_FILTER_OPTIONS["source_language"],
              interactive=True)
          target_language = gr.Dropdown(
              choices=SUPPORTED_TRANSLATION_LANGUAGES + [ANY_LANGUAGE],
              label="Target language",
              value=DEFAULT_FILTER_OPTIONS["target_language"],
              interactive=True)

        with gr.Row():
          filtered_translation = gr.DataFrame(
              headers=["Rank", "Model", "Elo rating"],
              datatype=["number", "str", "number"],
              value=lambda: load_elo_ratings(
                  LeaderboardTab.TRANSLATION, DEFAULT_FILTER_OPTIONS[
                      "source_language"], DEFAULT_FILTER_OPTIONS[
                          "target_language"]),
              elem_classes="leaderboard")

          source_language.change(fn=update_filtered_leaderboard,
                                 inputs=[
                                     gr.State(LeaderboardTab.TRANSLATION),
                                     gr.State(), source_language,
                                     target_language
                                 ],
                                 outputs=filtered_translation)
          target_language.change(fn=update_filtered_leaderboard,
                                 inputs=[
                                     gr.State(LeaderboardTab.TRANSLATION),
                                     gr.State(), source_language,
                                     target_language
                                 ],
                                 outputs=filtered_translation)

      # When filter options are changed, the accordion keeps closed.
      # To avoid this, we open the accordion when the filter options are changed.
      summary_language.change(fn=lambda: gr.Accordion(open=True),
                              outputs=summarization_filter)
      source_language.change(fn=lambda: gr.Accordion(open=True),
                             outputs=translation_filter)
      target_language.change(fn=lambda: gr.Accordion(open=True),
                             outputs=translation_filter)

      gr.Dataframe(headers=["Rank", "Model", "Elo rating"],
                   datatype=["number", "str", "number"],
                   value=lambda: load_elo_ratings(LeaderboardTab.TRANSLATION),
                   every=LEADERBOARD_UPDATE_INTERVAL,
                   elem_classes="leaderboard")
      gr.Markdown(LEADERBOARD_INFO)