-
Notifications
You must be signed in to change notification settings - Fork 1
/
evaluate_conditions.py
185 lines (155 loc) · 7.25 KB
/
evaluate_conditions.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
import os
import random
import csv
import tqdm
import argparse
from evaluate_llm import EvaluateLLM
from evaluate_llm import parse_chat_response
from LM_hf import *
import ipdb
import json
from utils import *
DATA_DIR = 'data'
CONDITION_DIR = os.path.join(DATA_DIR, 'conditions')
RESULTS_DIR = os.path.join(DATA_DIR, 'results/full')
# PROMPT_DIR = '../prompt_instructions'
random.seed(0)
def parse_answer(parsed_answer):
search_for = 'Answer: '
# Finding the position where 'Answer: ' ends
occ = parsed_answer.rfind(search_for)
if occ >= 0:
return parsed_answer[occ + len(search_for):].strip().rstrip('</s>')
else:
return 'Not found.'
def evaluate_condition(model_name, temperature, method,
init_belief, variable, condition, num_probs,
max_tokens, verbose, mcq, offset, args):
with open("./lm_paths.json", "r") as lm_paths:
paths = json.load(lm_paths)
llm = LM_nnsight(model_path=paths[model_name])
test_model = EvaluateLLM(llm, method=method)
csv_name = os.path.join(CONDITION_DIR, f'{init_belief}_{variable}_{condition}/stories.csv')
with open(csv_name, "r") as f:
reader = csv.reader(f, delimiter=";")
condition_rows = list(reader)
stories = []
questions = []
predicted_answers_parsed = []
flags_correct = []
flags_invalid = []
answer_keys = []
thoughts = []
right = 0
wrong = 0
anomaly = 0
tot = num_probs - offset
idx = 0
# Load intervention dict
if 'interv' in test_model.method:
test_model.load_interv(args)
for row in tqdm.tqdm(condition_rows[offset:num_probs]):
idx += 1
story = row[0]
question_orig = row[1]
question = row[1]
true_answer, wrong_answer = row[2], row[3]
answers = [true_answer, wrong_answer]
# ipdb.set_trace()
random.shuffle(answers)
if mcq:
question = f"{question}\nChoose one of the following:\na) {answers[0]}\nb) {answers[1]}"
predicted_answer, thought = test_model.predict_answer(story, question, args)
if answers[0] == true_answer:
answer_key = 'a)'
negative_answer_key = 'b)'
else:
answer_key = 'b)'
negative_answer_key = 'a)'
if mcq:
predicted_answer_parsed = parse_answer(predicted_answer)
correct = (predicted_answer_parsed[:2].lower()==answer_key)
incorrect = (predicted_answer_parsed[:2].lower()==negative_answer_key)
not_found = (not correct) and (not incorrect)
if not_found:
# Double check the answer.
choose_right = (answer_key in predicted_answer_parsed) or (true_answer.lower().rstrip('.') in predicted_answer_parsed.lower())
choose_wrong = (negative_answer_key in predicted_answer_parsed) or (wrong_answer.lower().rstrip('.') in predicted_answer_parsed.lower())
if choose_right and (not choose_wrong):
not_found = 0
correct = 1
if choose_wrong and (not choose_right):
not_found = 0
incorrect = 1
correct, not_found = int(correct), int(not_found)
right += correct
wrong += incorrect
anomaly += not_found
if verbose:
print_colored(f"THOUGHT: {thought}", "blue")
print_colored(f"PREDICT: {predicted_answer_parsed}", "green")
print(f"RIGHT: {answer_key} {true_answer}")
print(f"WRONG: {negative_answer_key} {wrong_answer}")
print_colored(f"GRADE: Right={correct}, Invalid={not_found}", "yellow")
print_colored(f"Right: Wrong: Anomaly = ({right}: {wrong}: {anomaly}) / {idx}. Acc: {right / idx:.2f}", "red")
stories.append(story)
questions.append(question)
predicted_answers_parsed.append(predicted_answer_parsed)
flags_correct.append(correct)
flags_invalid.append(not_found)
answer_keys.append(answer_key)
thoughts.append(thought)
# save results
model_name = model_name.replace('/', '_')
prediction = os.path.join(RESULTS_DIR, f'{init_belief}_{variable}_{condition}/Dir_{args.direction}_{args.dynamic}_{args.dvariable}/prediction_{model_name}_{temperature}_{method}_{variable}_{condition}_{offset}_{num_probs}.csv')
if not os.path.exists(os.path.join(RESULTS_DIR, f'{init_belief}_{variable}_{condition}/Dir_{args.direction}_{args.dynamic}_{args.dvariable}/')):
os.makedirs(os.path.join(RESULTS_DIR, f'{init_belief}_{variable}_{condition}/Dir_{args.direction}_{args.dynamic}_{args.dvariable}/'))
combined_results = zip(stories, questions, answer_keys, predicted_answers_parsed, flags_correct, flags_invalid, thoughts)
with open(prediction, "w") as f:
writer = csv.writer(f, delimiter=";")
# Write a header row
writer.writerow(["Story", "Question", "Correct Answer", "Predicted Answer", "Correct", "Invalid", "Thought"])
for row in combined_results:
writer.writerow(row)
accuracy = right / len(flags_correct)
# Print results
print("\n------------------------")
print(" RESULTS ")
print("------------------------")
print(f"MODEL: {model_name}, Temperature: {temperature}, Method: {method}")
print(f"CONDITION: {init_belief} {variable}, {condition}")
print(f"ACCURACY: {accuracy:.2%}")
print("------------------------\n")
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--variable', type=str, default='belief')
parser.add_argument('--dvariable', type=str, default='None')
parser.add_argument('--condition', type=str, default='true_belief')
parser.add_argument('--model_name', type=str, default='openai/text-davinci-003')
parser.add_argument('--temperature', type=float, default=0.0)
parser.add_argument('--num_probs', '-n', type=int, default=1)
parser.add_argument('--offset', '-o', type=int, default=0)
parser.add_argument('--max_tokens', type=int, default=100)
parser.add_argument('--method', type=str, default='0shot')
parser.add_argument('--init_belief', type=str, default="0_backward")
parser.add_argument('--verbose', '-v', action='store_true')
parser.add_argument('--mcq', action='store_true')
parser.add_argument('--belief', type=str, default='protagonist')
parser.add_argument('--dynamic', type=str, default='0_forward')
parser.add_argument('--K', type=int, default=32)
parser.add_argument('--alpha', type=int, default=20)
parser.add_argument('--direction', type=str, default="CoM")
parser.add_argument('--frac', type=float, default=0.)
args = parser.parse_args()
print("\nParameters:")
for attr, value in sorted(args.__dict__.items()):
print("\t{}={}".format(attr.upper(), value))
if args.dvariable=='None':
args.dvariable = args.variable
if 'interv' in args.method:
args.method += "_%s_K%d_a%d_%s" % (args.belief, args.K, args.alpha, args.direction)
evaluate_condition(args.model_name, args.temperature,
args.method, args.init_belief, args.variable,
args.condition, args.num_probs, args.max_tokens, args.verbose, args.mcq, args.offset, args)
if __name__ == '__main__':
main()