File size: 5,383 Bytes
a19f11e
 
 
 
 
 
 
 
076f69b
 
 
65566f3
a19f11e
 
 
50a5912
076f69b
 
50a5912
076f69b
 
65566f3
 
 
 
a19f11e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
65566f3
076f69b
a19f11e
 
076f69b
65566f3
 
 
 
 
 
 
 
 
a19f11e
65566f3
a19f11e
65566f3
 
 
a19f11e
 
 
 
 
 
 
 
 
 
519572d
 
 
a19f11e
 
 
 
 
 
 
 
65566f3
 
076f69b
65566f3
 
 
 
076f69b
65566f3
 
 
 
 
 
 
076f69b
 
65566f3
 
 
 
076f69b
 
 
a19f11e
076f69b
a19f11e
 
65566f3
b272a27
 
076f69b
a19f11e
076f69b
65566f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
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 pandas as pd

from credentials import get_credentials_json

# TODO(#21): Fix auto-reload issue related to the initialization of Firebase.
firebase_admin.initialize_app(credentials.Certificate(get_credentials_json()))
db = firestore.client()

SUPPORTED_TRANSLATION_LANGUAGES = [
    "Korean", "English", "Chinese", "Japanese", "Spanish", "French"
]


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

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

    if source_lang:
      collection = collection.where(filter=base_query.FieldFilter(
          "source_language", "==", source_lang.lower()))

    if target_lang:
      collection = collection.where(filter=base_query.FieldFilter(
          "target_language", "==", target_lang.lower()))

    return collection.stream()


def load_elo_ratings(tab, source_lang: str = None, target_lang: str = None):
  docs = get_docs(tab, 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)
  return [[i + 1, model, math.floor(rating + 0.5)]
          for i, (model, rating) in enumerate(sorted_ratings)]


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

DEFAULT_FILTER_OPTIONS = {
    "source_language": "English",
    "target_language": "Spanish"
}

filtered_dataframe = 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")


def update_filtered_leaderboard(source_lang, target_lang):
  new_value = load_elo_ratings(LeaderboardTab.TRANSLATION, source_lang,
                               target_lang)
  return gr.update(value=new_value)


def build_leaderboard():
  with gr.Tabs():
    with gr.Tab(LeaderboardTab.SUMMARIZATION.value):
      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):
        with gr.Row():
          source_language = gr.Dropdown(
              choices=SUPPORTED_TRANSLATION_LANGUAGES,
              label="Source language",
              value=DEFAULT_FILTER_OPTIONS["source_language"],
              interactive=True)
          target_language = gr.Dropdown(
              choices=SUPPORTED_TRANSLATION_LANGUAGES,
              label="Target language",
              value=DEFAULT_FILTER_OPTIONS["target_language"],
              interactive=True)

          source_language.change(fn=update_filtered_leaderboard,
                                 inputs=[source_language, target_language],
                                 outputs=filtered_dataframe)
          target_language.change(fn=update_filtered_leaderboard,
                                 inputs=[source_language, target_language],
                                 outputs=filtered_dataframe)

        with gr.Row():
          filtered_dataframe.render()

      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)