-
Notifications
You must be signed in to change notification settings - Fork 14
/
calc_adversarial_results_calc_score_calibration.py
302 lines (220 loc) · 11.3 KB
/
calc_adversarial_results_calc_score_calibration.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
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
import os
import argparse
import json
import pandas as pd
import numpy as np
def get_accuracy_calibration(human_results,
model_results,
model_results_reverse,
method="single",
example_nums=100, ):
assert method in ["single", "majority_vote", "average", "adversarial"]
predictions = []
targets = []
meta_results = {}
if method == "single":
for model_result, model_result_reverse in zip(model_results[:example_nums], model_results_reverse[:example_nums]):
try:
assist_one_score = float(
model_result["evaluation"][0]['evaluation'].split("\n")[-2].split("The score of Assistant 1:")[-1])
assist_two_score = float(
model_result["evaluation"][0]['evaluation'].split("\n")[-1].split("The score of Assistant 2:")[-1])
assist_one_score_reverse = float(
model_result_reverse["evaluation"][0]['evaluation'].split("\n")[-2].split("The score of Assistant 1:")[-1])
assist_two_score_reverse = float(
model_result_reverse["evaluation"][0]['evaluation'].split("\n")[-1].split("The score of Assistant 2:")[-1])
output_1_score = assist_one_score + assist_two_score_reverse
output_2_score = assist_two_score + assist_one_score_reverse
if output_1_score > output_2_score:
predictions.append("gpt35")
elif output_1_score < output_2_score:
predictions.append("vicuna")
elif output_1_score == output_2_score:
predictions.append("tie")
except BaseException as e:
predictions.append("tie")
for human_result in human_results[:example_nums]:
targets.append(human_result["human_results"])
correct = 0
for prediction, target in zip(predictions, targets):
if prediction == target:
correct += 1
meta_results["targets"] = targets
meta_results["predictions"] = predictions
return correct / len(targets), meta_results
elif method == "majority_vote":
for model_result in model_results[:example_nums]:
to_vote = []
for role in model_result["evaluation"]:
prediction = role['evaluation'].split("\n")[-1].split("Answer: ")[-1]
try:
if prediction == "Assistant 1":
to_vote.append("gpt35")
elif prediction == "Assistant 2":
to_vote.append("vicuna")
elif prediction == "Tie":
to_vote.append("tie")
else:
raise ValueError("check")
except BaseException as e:
to_vote.append("tie")
gpt35_num, vicuna_num, tie_num = 0, 0, 0
for ans in to_vote:
if ans == "gpt35":
gpt35_num += 1
elif ans == "vicuna":
vicuna_num += 1
elif ans == "tie":
tie_num += 1
dummy = [gpt35_num, vicuna_num, tie_num]
if len(set(dummy)) == 1:
predictions.append("tie")
elif dummy.index(max(dummy)) == 0:
predictions.append("gpt35")
elif dummy.index(max(dummy)) == 1:
predictions.append("vicuna")
elif dummy.index(max(dummy)) == 2:
predictions.append("tie")
for human_result in human_results[:example_nums]:
targets.append(human_result["human_results"])
correct = 0
for _, (prediction, target) in enumerate(zip(predictions, targets)):
if prediction == target:
correct += 1
meta_results["targets"] = targets
meta_results["predictions"] = predictions
return correct / len(targets), meta_results
elif method == "average":
for model_result, model_result_reverse in zip(model_results[:example_nums], model_results_reverse[:example_nums]):
output_1_score = 0
output_2_score = 0
try:
for model_role, model_role_reverse in zip(model_result["evaluation"], model_result_reverse["evaluation"]):
output_1_score += float(
model_role['evaluation'].split("\n")[-2].split("The score of Assistant 1:")[-1])
output_2_score += float(
model_role['evaluation'].split("\n")[-1].split("The score of Assistant 2:")[-1])
output_2_score += float(
model_role_reverse['evaluation'].split("\n")[-2].split("The score of Assistant 1:")[-1])
output_1_score += float(
model_role_reverse['evaluation'].split("\n")[-1].split("The score of Assistant 2:")[-1])
if output_1_score > output_2_score:
predictions.append("output_1")
elif output_1_score < output_2_score:
predictions.append("output_2")
elif output_1_score == output_2_score:
predictions.append("output_1")
except BaseException as e:
predictions.append("output_1")
for human_result in human_results[:example_nums]:
targets.append(human_result["human_results"])
correct = 0
for prediction, target in zip(predictions, targets):
if prediction == target:
correct += 1
meta_results["targets"] = targets
meta_results["predictions"] = predictions
return correct / len(targets), meta_results
elif method == "adversarial":
predictions_1 = []
predictions_2 = []
for model_result, model_result_reverse in zip(model_results[:example_nums], model_results_reverse[:example_nums]):
output_1_score = 0
output_2_score = 0
output_2_score_reverse = 0
output_1_score_reverse = 0
try:
for model_role in model_result["evaluation"]:
output_1_score += float(
model_role['evaluation'].split("\n")[-2].split("The score of Assistant 1:")[-1])
output_2_score += float(
model_role['evaluation'].split("\n")[-1].split("The score of Assistant 2:")[-1])
for model_role_reverse in model_result_reverse["evaluation"]:
output_2_score_reverse += float(
model_role_reverse['evaluation'].split("\n")[-2].split("The score of Assistant 1:")[-1])
output_1_score_reverse += float(
model_role_reverse['evaluation'].split("\n")[-1].split("The score of Assistant 2:")[-1])
if output_1_score > output_2_score:
predictions_1.append("output_1")
elif output_1_score < output_2_score:
predictions_1.append("output_2")
elif output_1_score == output_2_score:
predictions_1.append("output_1")
if output_2_score_reverse > output_1_score_reverse:
predictions_2.append("output_2")
elif output_2_score_reverse < output_1_score_reverse:
predictions_2.append("output_1")
elif output_2_score_reverse == output_1_score_reverse:
predictions_2.append("output_1")
except BaseException as e:
predictions_1.append("output_1")
predictions_2.append("output_1")
predictions = predictions_1 + predictions_2
for human_result in human_results[:example_nums] * 2:
targets.append(human_result["human_results"])
correct = 0
for prediction, target in zip(predictions, targets):
if prediction == target:
correct += 1
meta_results["targets"] = targets
meta_results["predictions"] = predictions
return correct / len(targets), meta_results
def get_kappa(meta:dict):
targets = meta["targets"]
predictions = meta["predictions"]
confusion = np.zeros((3, 3))
name2index = {
"gpt35": 0,
"vicuna": 1,
"tie": 2
}
for predict, target in zip(predictions, targets):
confusion[name2index[predict]][name2index[target]] += 1
n = np.sum(confusion)
sum_po = 0
sum_pe = 0
for i in range(len(confusion[0])):
sum_po += confusion[i][i]
row = np.sum(confusion[i, :])
col = np.sum(confusion[:, i])
sum_pe += row * col
po = sum_po / n
pe = sum_pe / (n * n)
# print(po, pe)
return (po - pe) / (1 - pe)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
# # single
# parser.add_argument("--human_results_path", default="./agentverse/tasks/llm_eval/data/faireval/preprocessed_data"
# "/test.json")
# parser.add_argument("--model_results_path",
# default="./outputs/llm_eval/single_role/faireval/calc_score_comparison")
# parser.add_argument("--model_results_reverse_path",
# default="./outputs/llm_eval/single_role/faireval/calc_score_comparison_reverse")
# parser.add_argument("--output_path", default="./outputs/llm_eval/single_role/faireval/calc_score_comparison_calibration")
# parser.add_argument("--ensemble_method", default="single")
# # multi
parser.add_argument("--human_results_path", default="./agentverse/tasks/llm_eval/data/adversarial/preprocessed_data"
"/test.json")
parser.add_argument("--model_results_path",
default="./outputs/llm_eval/multi_role/only_static_assign/adversarial/two_turns_sequential/three_different_role/calc_score_comparison/gpt_35_0613")
parser.add_argument("--model_results_reverse_path",
default="./outputs/llm_eval/multi_role/only_static_assign/adversarial/two_turns_sequential/three_different_role/calc_score_comparison_reverse/gpt_35_0613")
parser.add_argument("--output_path", default="./outputs/llm_eval/multi_role/only_static_assign/adversarial/two_turns_sequential/three_different_role/calc_score_comparison_calibration")
parser.add_argument("--ensemble_method", default="average")
args = parser.parse_args()
with open(args.human_results_path) as f:
human_results = json.load(f)
with open(os.path.join(args.model_results_path, "pair_comparison_results.json")) as f:
model_results = json.load(f)
with open(os.path.join(args.model_results_reverse_path, "pair_comparison_results.json")) as f:
model_results_reverse = json.load(f)
accuracy, meta = get_accuracy_calibration(human_results=human_results,
model_results=model_results,
model_results_reverse=model_results_reverse,
method=args.ensemble_method,
example_nums=len(human_results),
)
# kappa = get_kappa(meta)
print(f"accuracy: {accuracy}")
# print(f"kappa: {kappa}")