-
Notifications
You must be signed in to change notification settings - Fork 60
/
gemini_runner.py
195 lines (172 loc) · 6.35 KB
/
gemini_runner.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
from concurrent.futures import ThreadPoolExecutor, as_completed
import os
from time import time
import pandas as pd
from tqdm import tqdm
from vertexai.preview.generative_models import GenerativeModel
from eval.eval import compare_query_results
from utils.creds import db_creds_all
from utils.pruning import prune_metadata_str
from utils.questions import prepare_questions_df
from utils.reporting import upload_results
def multiturn_generate_content(model_name="gemini-pro"):
config = {"max_output_tokens": 600, "temperature": 0, "top_p": 1}
model = GenerativeModel(model_name, generation_config=config)
chat = model.start_chat()
return chat
def generate_prompt(
prompt_file,
question,
db_name,
instructions="",
k_shot_prompt="",
glossary="",
table_metadata_string="",
prev_invalid_sql="",
prev_error_msg="",
public_data=True,
num_columns_to_keep=20,
shuffle=True,
):
if "gemini" not in prompt_file:
raise ValueError("Invalid prompt file. Please use prompt_gemini.md")
with open(prompt_file, "r") as f:
prompt = f.read()
question_instructions = question + " " + instructions
if table_metadata_string == "":
pruned_metadata_ddl, join_str = prune_metadata_str(
question_instructions, db_name, public_data, num_columns_to_keep, shuffle
)
pruned_metadata_str = pruned_metadata_ddl + join_str
else:
pruned_metadata_str = table_metadata_string
prompt = prompt.format(
user_question=question,
instructions=instructions,
table_metadata_string=pruned_metadata_str,
k_shot_prompt=k_shot_prompt,
glossary=glossary,
prev_invalid_sql=prev_invalid_sql,
prev_error_msg=prev_error_msg,
)
return prompt
def process_row(row, model_name, args):
start_time = time()
chat = multiturn_generate_content(model_name=model_name)
response = chat.send_message(row["prompt"])
end_time = time()
generated_query = response.text.split("```sql")[-1].split("```")[0].strip()
row["generated_query"] = generated_query
row["latency_seconds"] = end_time - start_time
golden_query = row["query"]
db_name = row["db_name"]
db_type = row["db_type"]
question = row["question"]
query_category = row["query_category"]
exact_match = correct = 0
try:
exact_match, correct = compare_query_results(
query_gold=golden_query,
query_gen=generated_query,
db_name=db_name,
db_type=db_type,
db_creds=db_creds_all[row["db_type"]],
question=question,
query_category=query_category,
decimal_points=args.decimal_points,
)
row["exact_match"] = int(exact_match)
row["correct"] = int(correct)
row["error_msg"] = ""
except Exception as e:
row["error_db_exec"] = 1
row["error_msg"] = f"QUERY EXECUTION ERROR: {e}"
return row
def run_gemini_eval(args):
# get params from args
questions_file_list = args.questions_file
prompt_file_list = args.prompt_file
num_questions = args.num_questions
public_data = not args.use_private_data
model_name = args.model
output_file_list = args.output_file
k_shot = args.k_shot
max_workers = args.parallel_threads
db_type = args.db_type
cot_table_alias = args.cot_table_alias
for questions_file, prompt_file, output_file in zip(
questions_file_list, prompt_file_list, output_file_list
):
print(f"Using prompt file {prompt_file}")
# get questions
print("Preparing questions...")
print(
f"Using {'all' if num_questions is None else num_questions} question(s) from {questions_file}"
)
df = prepare_questions_df(
questions_file, db_type, num_questions, k_shot, cot_table_alias
)
# create a prompt for each question
df["prompt"] = df.apply(
lambda row: generate_prompt(
prompt_file,
row["question"],
row["db_name"],
row["db_type"],
row["instructions"],
row["k_shot_prompt"],
row["glossary"],
row["table_metadata_string"],
row["prev_invalid_sql"],
row["prev_error_msg"],
row["cot_instructions"],
row["cot_pregen"],
public_data,
args.num_columns,
args.shuffle_metadata,
),
axis=1,
)
total_tried = 0
total_correct = 0
output_rows = []
print(f"Running evaluation using {model_name}...")
with ThreadPoolExecutor(max_workers=max_workers) as executor:
futures = []
for row in df.to_dict("records"):
futures.append(executor.submit(process_row, row, model_name, args))
with tqdm(as_completed(futures), total=len(futures)) as pbar:
for f in pbar:
row = f.result()
output_rows.append(row)
if row["correct"]:
total_correct += 1
total_tried += 1
pbar.update(1)
pbar.set_description(
f"Correct so far: {total_correct}/{total_tried} ({100*total_correct/total_tried:.2f}%)"
)
output_df = pd.DataFrame(output_rows)
del output_df["prompt"]
print(output_df.groupby("query_category")[["correct", "error_db_exec"]].mean())
output_df = output_df.sort_values(by=["db_name", "query_category", "question"])
# get directory of output_file and create if not exist
output_dir = os.path.dirname(output_file)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
try:
output_df.to_csv(output_file, index=False, float_format="%.2f")
except:
output_df.to_pickle(output_file)
results = output_df.to_dict("records")
# upload results
with open(prompt_file, "r") as f:
prompt = f.read()
if args.upload_url is not None:
upload_results(
results=results,
url=args.upload_url,
runner_type="api_runner",
prompt=prompt,
args=args,
)