forked from lmarena/arena-hard-auto
-
Notifications
You must be signed in to change notification settings - Fork 0
/
show_result.py
276 lines (223 loc) · 10.6 KB
/
show_result.py
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
import pandas as pd
import numpy as np
import plotly.express as px
import datetime
import argparse
import os
import math
from glob import glob
from tqdm import tqdm
import inspect
from sklearn.linear_model import LogisticRegression
from collections import defaultdict
from utils import load_model_answers
def compute_mle_elo(df, SCALE=400, BASE=10, INIT_RATING=1000, baseline_model="gpt-4-0314"):
models = pd.concat([df["model_a"], df["model_b"]]).unique()
models = pd.Series(np.arange(len(models)), index=models)
# duplicate battles
df = pd.concat([df, df], ignore_index=True)
p = len(models.index)
n = df.shape[0]
X = np.zeros([n, p])
X[np.arange(n), models[df["model_a"]]] = +math.log(BASE)
X[np.arange(n), models[df["model_b"]]] = -math.log(BASE)
# one A win => two A win
Y = np.zeros(n)
Y[df["winner"] == "model_a"] = 1.0
# one tie => one A win + one B win
# find tie + tie (both bad) index
tie_idx = (df["winner"] == "tie") | (df["winner"] == "tie (bothbad)")
tie_idx[len(tie_idx)//2:] = False
Y[tie_idx] = 1.0
lr = LogisticRegression(fit_intercept=False, penalty=None, tol=1e-8)
lr.fit(X,Y)
elo_scores = SCALE * lr.coef_[0] + INIT_RATING
# set anchor as gpt-4-0314 = 1000
if baseline_model in models.index:
elo_scores += 1000 - elo_scores[models[baseline_model]]
return pd.Series(elo_scores, index=models.index).sort_values(ascending=False)
def get_bootstrap_result(battles, func_compute_elo, num_round, baseline_model="gpt-4-0314"):
rows = []
kwargs = {}
if baseline_model in inspect.signature(func_compute_elo).parameters:
kwargs[baseline_model] = baseline_model
for _ in tqdm(range(num_round), desc="bootstrap"):
rows.append(func_compute_elo(battles.sample(frac=1.0, replace=True), **kwargs))
df = pd.DataFrame(rows)
return df[df.median().sort_values(ascending=False).index]
def preety_print_two_ratings(ratings_1, ratings_2, column_names):
df = pd.DataFrame([
[n, ratings_1[n], ratings_2[n]] for n in ratings_1.keys()
], columns=["Model", column_names[0], column_names[1]]).sort_values(column_names[0], ascending=False).reset_index(drop=True)
df[column_names[0]] = (df[column_names[0]] + 0.5).astype(int)
df[column_names[1]] = (df[column_names[1]] + 0.5).astype(int)
df.index = df.index + 1
return df
def visualize_bootstrap_scores(df, title):
bars = pd.DataFrame(dict(
lower = df.quantile(.025),
rating = df.quantile(.5),
upper = df.quantile(.975))).reset_index(names="model").sort_values("rating", ascending=False)
bars['error_y'] = bars['upper'] - bars["rating"]
bars['error_y_minus'] = bars['rating'] - bars["lower"]
bars['rating_rounded'] = np.round(bars['rating'], 2)
fig = px.scatter(bars, x="model", y="rating", error_y="error_y",
error_y_minus="error_y_minus", text="rating_rounded",
title=title)
fig.update_layout(xaxis_title="Model", yaxis_title="Rating",
height=600)
return fig
def predict_win_rate(elo_ratings, SCALE=400, BASE=10, INIT_RATING=1000):
names = sorted(list(elo_ratings.keys()))
wins = defaultdict(lambda: defaultdict(lambda: 0))
for a in names:
for b in names:
ea = 1 / (1 + BASE ** ((elo_ratings[b] - elo_ratings[a]) / SCALE))
wins[a][b] = ea
wins[b][a] = 1 - ea
data = {
a: [wins[a][b] if a != b else np.NAN for b in names]
for a in names
}
df = pd.DataFrame(data, index=names)
df.index.name = "model_a"
df.columns.name = "model_b"
return df.T
def get_win_rate_column(df, column, baseline="gpt-4-0314"):
to_dict = df[["model", column]].set_index("model").to_dict()[column]
win_rate_table = predict_win_rate(to_dict)
return win_rate_table[baseline].fillna(0.5).apply(lambda x: round(x * 100, 2))
def get_battles_from_judgment(judge_name, first_game_only=False, WEIGHT=3, baseline_model="gpt-4-0314"):
arena_hard_battles = pd.DataFrame()
print("Turning judgment results into battles...")
directory = f"data/arena-hard-v0.1/model_judgment/{judge_name}"
assert os.path.exists(directory)
for file in tqdm(glob(f"{directory}/*jsonl")):
df = pd.read_json(file, lines=True)
for _, row in df.iterrows():
# game 1
output = {"question_id": row["question_id"],
"model_a": baseline_model,
"model_b": row["model"]}
game = row["games"][0]
weight = 1
if game["score"] == "A=B":
output["winner"] = "tie"
elif game["score"] == "A>B":
output["winner"] = "model_a"
elif game["score"] == "A>>B":
output["winner"] = "model_a"
weight = WEIGHT
elif game["score"] == "B>A":
output["winner"] = "model_b"
elif game["score"] == "B>>A":
output["winner"] = "model_b"
weight = WEIGHT
else:
weight = 0
if weight:
arena_hard_battles = pd.concat([arena_hard_battles, pd.DataFrame([output] * weight)])
if not first_game_only:
# game 2
output = {"question_id": row["question_id"],
"model_a": baseline_model,
"model_b": row["model"]}
game = row["games"][1]
weight = 1
if game["score"] == "A=B":
output["winner"] = "tie"
elif game["score"] == "A>B":
output["winner"] = "model_b"
elif game["score"] == "A>>B":
output["winner"] = "model_b"
weight = WEIGHT
elif game["score"] == "B>A":
output["winner"] = "model_a"
elif game["score"] == "B>>A":
output["winner"] = "model_a"
weight = WEIGHT
else:
weight = 0
if weight:
arena_hard_battles = pd.concat([arena_hard_battles, pd.DataFrame([output] * weight)])
arena_hard_battles.to_json("data/arena_hard_battles.jsonl", lines=True, orient="records")
return arena_hard_battles
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--bench-name", type=str, default="arena-hard-v0.1")
parser.add_argument("--judge-name", type=str, default="gpt-4-1106-preview")
parser.add_argument("--baseline", type=str, default="gpt-4-0314")
parser.add_argument("--load-battles", action="store_true")
parser.add_argument("--load-bootstrap", action="store_true")
parser.add_argument("--show-elo", action="store_true")
parser.add_argument("--weight", type=int, default=3)
parser.add_argument("--num-rounds", type=int, default=100)
parser.add_argument("--output", action="store_true")
parser.add_argument("--first-game-only", action="store_true")
args = parser.parse_args()
print(args)
assert not args.load_bootstrap or (args.load_battles and args.load_bootstrap), "If loading prexisting bootstrapping data, you must also load preexisting battles."
answer_dir = os.path.join("data", args.bench_name, "model_answer")
model_answers = load_model_answers(answer_dir)
if args.load_battles:
assert os.path.exists("data/arena_hard_battles.jsonl")
battles = pd.read_json("data/arena_hard_battles.jsonl", lines=True)
else:
battles = get_battles_from_judgment(args.judge_name, args.first_game_only, args.weight, args.baseline)
bootstrap_online_elo = compute_mle_elo(battles, baseline_model=args.baseline)
if args.load_bootstrap:
bootstrap_elo_lu = pd.read_json("data/bootstrapping_results.jsonl", lines=True)
else:
np.random.seed(42)
bootstrap_elo_lu = get_bootstrap_result(battles, compute_mle_elo, args.num_rounds, args.baseline)
bootstrap_elo_lu.to_json("data/bootstrapping_results.jsonl", lines=True, orient="records")
stats = pd.DataFrame()
stats["results"] = None
stats["results"] = stats['results'].astype('object')
for i, model in enumerate(bootstrap_online_elo.index):
assert model in bootstrap_elo_lu.columns
stats.at[i, "model"] = model
stats.at[i, "score"] = bootstrap_online_elo[model]
stats.at[i, "lower"] = np.percentile(bootstrap_elo_lu[model], 2.5)
stats.at[i, "upper"] = np.percentile(bootstrap_elo_lu[model], 97.5)
length = 0
if model in model_answers:
for _, row in model_answers[model].items():
turn = row["choices"][0]["turns"][0]
length += turn["token_len"]
length /= len(model_answers[model])
stats.at[i, "avg_tokens"] = int(length)
stats.at[i, "results"] = bootstrap_elo_lu[model].tolist()
if not args.show_elo:
stats.sort_values(by="model", inplace=True)
stats["score"] = get_win_rate_column(stats, "score", args.baseline).tolist()
stats["lower"] = get_win_rate_column(stats, "lower", args.baseline).tolist()
stats["upper"] = get_win_rate_column(stats, "upper", args.baseline).tolist()
decimal = 1
else:
decimal = 0
stats = stats.astype({"score" : int, "lower" : int, "upper" : int})
stats.sort_values(by="score", ascending=False, inplace=True)
for _, row in stats.iterrows():
interval = str((round(row['lower'] - row['score'], decimal), round(row['upper'] - row['score'], decimal)))
print(f"{row['model'] : <30} | score: {round(row['score'], decimal) : ^5} | 95% CI: {interval : ^12} | average #tokens: {int(row['avg_tokens'])}")
if args.output:
cur_date = datetime.datetime.now()
date_str = cur_date.strftime("%Y%m%d")
stats = stats.drop(columns=['results'])
CI = []
for i in range(len(stats)):
score = stats.iloc[i]['score']
upper = stats.iloc[i]['upper']
lower = stats.iloc[i]['lower']
CI.append(f"(-{(score-lower):.2f}, +{(upper-score):.2f})")
stats["CI"] = CI
col_list = list(stats)
stats = stats.loc[:,col_list]
stats.rename(columns={'upper': 'rating_q975'}, inplace=True)
stats.rename(columns={'lower': 'rating_q025'}, inplace=True)
col_list = list(stats)
col_list[-2], col_list[-1] = col_list[-1], col_list[-2]
stats = stats.loc[:,col_list]
stats['date'] = date_str[:4] + '-' + date_str[4:6] + '-' + date_str[6:]
stats.to_csv(f"leaderboard/arena_hard_leaderboard_{date_str}.csv", index=False)