Skip to content

armgit / camel

Public Imported from GitHub
0

🐫 CAMEL: The first and the best multi-agent framework. Finding the Scaling Law of Agents. https://www.camel-ai.org

Fix data generation (#27)

Hasan Abed Al Kader Hammoud001f92edb1a7
diff --git a/examples/role_playing_multiprocess.py b/examples/role_playing_multiprocess.py
index 873bebd49..f406a8d3d 100644
--- a/examples/role_playing_multiprocess.py
+++ b/examples/role_playing_multiprocess.py
@@ -42,7 +42,7 @@ def generate_data(assistant_idx: int, assistant_role_name: str, user_idx: int,
user_role_name: str, task_idx: int,
task_prompt: str) -> None:
- max_num_messages = 20
+ max_num_messages = 30
original_task_prompt = task_prompt
task_specify_agent = TaskSpecifyAgent(ModeType.GPT_3_5_TURBO,
@@ -80,13 +80,27 @@ def generate_data(assistant_idx: int, assistant_role_name: str, user_idx: int,
# Threshold to terminate the conversation if no end token appears
repeat_word_counter = 0
- repeat_word_threshold = 3
- repeat_word_list = ["goodbye", "good bye", "thank", "bye", "welcome"]
+ repeat_word_threshold = 4
+ repeat_word_list = [
+ "goodbye", "good bye", "thank", "bye", "welcome",
+ "As an AI language model,"
+ ]
- # Set maximum number of messages for the chat
+ assistant_instruct_counter = 0
+ assistant_instruct_threshold = 1
+ assistant_instruct_word = "Instruction:"
+
+ user_no_instruct_counter = 0
+ user_no_instruct_threshold = 3
+ user_no_instruct_word = "Instruction:"
+
+ # Set max number of messages for the chat
while message_counter < max_num_messages:
+
user_msgs, user_terminated, user_info = user_agent.step(assistant_msg)
+
+ # Condition 1: User terminates the chat
if user_terminated:
message_dict["termination_reason"] = (
f"{str(user_agent.role_type)}: "
@@ -97,37 +111,66 @@ def generate_data(assistant_idx: int, assistant_role_name: str, user_idx: int,
print(f"User:\n{user_msg.content}\n")
user_msg.role = "user"
- message_counter += 1
- message_dict[f"message_{message_counter}"] = user_msg.to_dict()
-
- if "<CAMEL_TASK_DONE>" in user_msg.content:
- message_dict['termination_reason'] = "<CAMEL_TASK_DONE>"
- break
-
assistant_msgs, assistant_terminated, assistant_info = (
assistant_agent.step(user_msg))
+
+ # Condition 2: Assistant terminates the chat
if assistant_terminated:
message_dict["termination_reason"] = (
f"{str(assistant_agent.role_type)}: "
f"{assistant_info['finish_reasons'][0]}")
break
+
assistant_msg = assistant_msgs[0]
print(f"Assistant:\n{assistant_msg.content}\n")
assistant_msg.role = "user"
- message_counter += 1
- message_dict[f"message_{message_counter}"] = assistant_msg.to_dict()
-
+ # Condition 3: Break if user does not give instruction
+ if user_no_instruct_word in user_msg.content:
+ user_no_instruct_counter += 1
+ if user_no_instruct_counter == user_no_instruct_threshold:
+ message_dict[
+ 'termination_reason'] = "user_no_instruct_threshold"
+ break
+
+ # Condition 4: Break if assistant gives instruction (flipped role)
+ if assistant_instruct_word in assistant_msg.content:
+ assistant_instruct_counter += 1
+ if assistant_instruct_counter == assistant_instruct_threshold:
+ message_dict[
+ 'termination_reason'] = "assistant_instruct_threshold"
+ break
+
+ # Condition 5: Repeat word observed
for repeat_word in repeat_word_list:
- if repeat_word in user_msg.content.lower():
+ if repeat_word in user_msg.content.lower(
+ ) or repeat_word in assistant_msg.content.lower():
repeat_word_counter += 1
if repeat_word_counter == repeat_word_threshold:
+ message_dict[
+ 'termination_reason'] = "repeat_word_threshold"
break
else:
repeat_word_counter = 0
+ # Save user message
+ message_counter += 1
+ message_dict[f"message_{message_counter}"] = user_msg.to_dict()
+
+ # Condition 5: End token observed
+ if "<CAMEL_TASK_DONE>" in user_msg.content:
+ message_dict['termination_reason'] = "<CAMEL_TASK_DONE>"
+ break
+
+ # Save assistant message
+ message_counter += 1
+ message_dict[f"message_{message_counter}"] = assistant_msg.to_dict()
+
message_dict["num_messages"] = message_counter
+ if message_dict["num_messages"] == max_num_messages:
+ message_dict["termination_reason"] = "max_num_messages"
+
with open(f"./camel_data/{message_dict['id']}.json", "w") as json_file:
json.dump(message_dict, json_file)
@@ -135,13 +178,12 @@ def generate_data(assistant_idx: int, assistant_role_name: str, user_idx: int,
def main() -> None:
# Chunk for parallel jobs
- # array_idx = int(os.environ.get('SLURM_ARRAY_TASK_ID'))
- array_idx = 1
+ array_idx = int(os.environ.get('SLURM_ARRAY_TASK_ID'))
roles_per_chunk = 10
# Parameters for filtering the generated task string
start_token = "1."
- num_tasks = 40
+ num_tasks = 10
with open("./data/user_roles.txt", "r") as f:
user_roles = f.read().splitlines()
@@ -157,8 +199,10 @@ def main() -> None:
pool = multiprocessing.Pool()
for assistant_idx, assistant_role_name in enumerate(assistant_roles):
+ assistant_idx += array_idx * roles_per_chunk
+ assistant_role_name = " ".join(assistant_role_name.split(" ")[1:])
for user_idx, user_role_name in enumerate(user_roles):
-
+ user_role_name = " ".join(user_role_name.split(" ")[1:])
# Load the task list assigned for assistant and user roles
with open(f"./tasks/{assistant_role_name}_{user_role_name}.txt",
"r") as f: