forked from TIGER-AI-Lab/MMLU-Pro
-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluate_from_api.py
333 lines (298 loc) · 11.3 KB
/
evaluate_from_api.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
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
import os
from openai import AzureOpenAI
import openai
from openai import OpenAI
import anthropic
import google.generativeai as genai
import json
import re
import random
from tqdm import tqdm
import time
from datasets import load_dataset
import argparse
API_KEY = "Put your api key here"
def get_client():
if args.model_name in ["gpt-4", "gpt-4o"]:
openai.api_key = API_KEY
client = openai
elif args.model_name in ["deepseek-chat", "deepseek-coder"]:
client = OpenAI(api_key=API_KEY, base_url="https://api.deepseek.com/")
elif args.model_name in ["gemini-1.5-flash-latest", "gemini-1.5-pro-latest"]:
genai.configure(api_key=API_KEY)
generation_config = {
"temperature": 0.0,
"top_p": 1,
"max_output_tokens": 4000,
"response_mime_type": "text/plain",
}
safety_settings = [
{
"category": "HARM_CATEGORY_HARASSMENT",
"threshold": "BLOCK_NONE",
},
{
"category": "HARM_CATEGORY_HATE_SPEECH",
"threshold": "BLOCK_NONE",
},
{
"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
"threshold": "BLOCK_NONE",
},
{
"category": "HARM_CATEGORY_DANGEROUS_CONTENT",
"threshold": "BLOCK_NONE",
},
]
client = genai.GenerativeModel(
model_name=args.model_name,
safety_settings=safety_settings,
generation_config=generation_config,
)
elif args.model_name in ["claude-3-opus-20240229", "claude-3-sonnet-20240229"]:
client = anthropic.Anthropic(
api_key=API_KEY,
)
else:
client = None
print("For other model API calls, please implement the client definition method yourself.")
return client
def call_api(client, instruction, inputs):
start = time.time()
if args.model_name in ["gpt-4", "gpt-4o", "deepseek-chat", "deepseek-coder"]:
message_text = [{"role": "user", "content": instruction + inputs}]
completion = client.chat.completions.create(
model=args.model_name,
messages=message_text,
temperature=0,
max_tokens=4000,
top_p=1,
frequency_penalty=0,
presence_penalty=0,
stop=None
)
result = completion.choices[0].message.content
elif args.model_name in ["gemini-1.5-flash-latest", "gemini-1.5-pro-latest"]:
chat_session = client.start_chat(
history=[]
)
result = chat_session.send_message(instruction + inputs).text
elif args.model_name in ["claude-3-opus-20240229", "claude-3-sonnet-20240229"]:
message = client.messages.create(
model=args.model_name,
max_tokens=4000,
system="",
messages=[
{"role": "user", "content": instruction + inputs}
],
temperature=0.0,
top_p=1,
)
result = message.content[0]
else:
print("For other model API calls, please implement the request method yourself.")
result = None
print("cost time", time.time() - start)
return result
def load_mmlu_pro(dataset_id):
dataset = load_dataset(dataset_id)
test_df, val_df = dataset["test"], dataset["validation"]
test_df = preprocess(test_df)
val_df = preprocess(val_df)
return test_df, val_df
def preprocess(test_df):
res_df = []
for each in test_df:
options = []
for opt in each["options"]:
if opt == "N/A":
continue
options.append(opt)
each["options"] = options
res_df.append(each)
res = {}
for each in res_df:
if each["category"] not in res:
res[each["category"]] = []
res[each["category"]].append(each)
return res
def format_example(question, options, cot_content=""):
if cot_content == "":
cot_content = "Let's think step by step."
if cot_content.startswith("A: "):
cot_content = cot_content[3:]
example = "Question: {}\nOptions: ".format(question)
choice_map = "ABCDEFGHIJ"
for i, opt in enumerate(options):
example += "{}. {}\n".format(choice_map[i], opt)
if cot_content == "":
example += "Answer: "
else:
example += "Answer: " + cot_content + "\n\n"
return example
def extract_answer(text):
pattern = r"answer is \(?([A-J])\)?"
match = re.search(pattern, text)
if match:
return match.group(1)
else:
print("1st answer extract failed\n" + text)
return extract_again(text)
def extract_again(text):
match = re.search(r'.*[aA]nswer:\s*([A-J])', text)
if match:
return match.group(1)
else:
return extract_final(text)
def extract_final(text):
pattern = r"[A-J](?=[^A-J]*$)"
match = re.search(pattern, text)
if match:
return match.group(0)
else:
return None
def single_request(client, single_question, cot_examples_dict, exist_result):
exist = True
q_id = single_question["question_id"]
for each in exist_result:
if q_id == each["question_id"] and single_question["question"] == each["question"]:
pred = extract_answer(each["model_outputs"])
return pred, each["model_outputs"], exist
exist = False
category = single_question["category"]
cot_examples = cot_examples_dict[category]
question = single_question["question"]
options = single_question["options"]
prompt = "The following are multiple choice questions (with answers) about {}. Think step by" \
" step and then output the answer in the format of \"The answer is (X)\" at the end.\n\n" \
.format(category)
for each in cot_examples:
prompt += format_example(each["question"], each["options"], each["cot_content"])
input_text = format_example(question, options)
try:
start = time.time()
response = call_api(client, prompt, input_text)
print("requesting gpt 4 costs: ", time.time() - start)
except Exception as e:
print("error", e)
return None, None, exist
pred = extract_answer(response)
return pred, response, exist
def update_result(output_res_path):
category_record = {}
res = []
success = False
while not success:
try:
if os.path.exists(output_res_path):
with open(output_res_path, "r") as fi:
res = json.load(fi)
for each in res:
category = each["category"]
if category not in category_record:
category_record[category] = {"corr": 0.0, "wrong": 0.0}
if not each["pred"]:
random.seed(12345)
x = random.randint(0, len(each["options"]) - 1)
if x == each["answer_index"]:
category_record[category]["corr"] += 1
# print("random hit.")
else:
category_record[category]["wrong"] += 1
elif each["pred"] == each["answer"]:
category_record[category]["corr"] += 1
else:
category_record[category]["wrong"] += 1
success = True
except Exception as e:
print("Error", e, "sleep 2 seconds")
time.sleep(2)
return res, category_record
def merge_result(res, curr):
merged = False
for i, single in enumerate(res):
if single["question_id"] == curr["question_id"] and single["question"] == curr["question"]:
res[i] = curr
merged = True
if not merged:
res.append(curr)
return res
def evaluate(subjects, dataset_id):
client = get_client()
test_df, dev_df = load_mmlu_pro(dataset_id)
if not subjects:
subjects = list(test_df.keys())
print("assigned subjects", subjects)
for subject in subjects:
test_data = test_df[subject]
output_res_path = os.path.join(args.output_dir, subject + "_result.json")
output_summary_path = os.path.join(args.output_dir, subject + "_summary.json")
res, category_record = update_result(output_res_path)
for each in tqdm(test_data):
label = each["answer"]
category = subject
pred, response, exist = single_request(client, each, dev_df, res)
# if exist:
# continue
if response is not None:
res, category_record = update_result(output_res_path)
if category not in category_record:
category_record[category] = {"corr": 0.0, "wrong": 0.0}
each["pred"] = pred
each["model_outputs"] = response
merge_result(res, each)
if pred is not None:
if pred == label:
category_record[category]["corr"] += 1
else:
category_record[category]["wrong"] += 1
else:
category_record[category]["wrong"] += 1
save_res(res, output_res_path)
save_summary(category_record, output_summary_path)
res, category_record = update_result(output_res_path)
save_res(res, output_res_path)
save_summary(category_record, output_summary_path)
def save_res(res, output_res_path):
temp = []
exist_q_id = []
for each in res:
if each["question_id"] not in exist_q_id:
exist_q_id.append(each["question_id"])
temp.append(each)
else:
continue
res = temp
with open(output_res_path, "w") as fo:
fo.write(json.dumps(res))
def save_summary(category_record, output_summary_path):
total_corr = 0.0
total_wrong = 0.0
for k, v in category_record.items():
if k == "total":
continue
cat_acc = v["corr"] / (v["corr"] + v["wrong"])
category_record[k]["acc"] = cat_acc
total_corr += v["corr"]
total_wrong += v["wrong"]
acc = total_corr / (total_corr + total_wrong)
category_record["total"] = {"corr": total_corr, "wrong": total_wrong, "acc": acc}
with open(output_summary_path, "w") as fo:
fo.write(json.dumps(category_record))
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--output_dir", "-o", type=str, default="eval_results/")
parser.add_argument("--model_name", "-m", type=str, default="gpt-4",
choices=["gpt-4", "gpt-4o", "deepseek-chat", "deepseek-coder",
"gemini-1.5-flash-latest", "gemini-1.5-pro-latest",
"claude-3-opus-20240229", "claude-3-sonnet-20240229"])
parser.add_argument("--assigned_subjects", "-a", type=str, default="all")
parser.add_argument("--dataset", "-d", type=str, default="TIGER-Lab/MMLU-Pro")
assigned_subjects = []
args = parser.parse_args()
if args.assigned_subjects == "all":
assigned_subjects = []
else:
assigned_subjects = args.assigned_subjects.split(",")
os.makedirs(args.output_dir, exist_ok=True)
evaluate(assigned_subjects, args.dataset)