When running the benchmark, we should now use random probes instead of "random" or "What do you do in London as a tourist?"
Browse files- run_benchmark.py +2 -2
run_benchmark.py
CHANGED
@@ -6,7 +6,7 @@ Use this file to test how a model does with the system prompts, with no addition
|
|
6 |
from openai import OpenAI
|
7 |
import configparser
|
8 |
from hundred_system_prompts import *
|
9 |
-
|
10 |
|
11 |
# Get config
|
12 |
config = configparser.ConfigParser()
|
@@ -23,7 +23,7 @@ with open(filename, "w") as file:
|
|
23 |
for i, (prompt, probe, lambda_function) in enumerate(system_prompts):
|
24 |
history = [
|
25 |
{"role": "system", "content": prompt},
|
26 |
-
{"role": "user", "content": probe}
|
27 |
]
|
28 |
|
29 |
completion = client.chat.completions.create(
|
|
|
6 |
from openai import OpenAI
|
7 |
import configparser
|
8 |
from hundred_system_prompts import *
|
9 |
+
import random
|
10 |
|
11 |
# Get config
|
12 |
config = configparser.ConfigParser()
|
|
|
23 |
for i, (prompt, probe, lambda_function) in enumerate(system_prompts):
|
24 |
history = [
|
25 |
{"role": "system", "content": prompt},
|
26 |
+
{"role": "user", "content": probe if probe != "random" else random.choice(random_probes)}
|
27 |
]
|
28 |
|
29 |
completion = client.chat.completions.create(
|