π The First Self-Improving agents with RL / Prompting Optimization
Python agentaigptmulti-agentpythonreinforcement-learningself-improvementsociety-of-mindswarm-intelligence
import pytest
from typing import Optional
from swarm.graph.swarm import Swarm
from swarm.environment.operations.final_decision import MergingStrategy
from experiments.evaluator.evaluator import Evaluator
from experiments.evaluator.datasets.mmlu_dataset import MMLUDataset
from datasets.MMLU.download import download
@pytest.fixture
def mmlu_data():
download()
return None
def make_swarm(model_name: Optional[str]) -> Swarm:
agent_name_list = 1 * ["IO"] + 1 * ["AdversarialAgent"]
domain = 'mmlu'
swarm = Swarm(
agent_name_list,
domain,
model_name=model_name,
final_node_class="FinalDecision",
final_node_kwargs=dict(strategy=MergingStrategy.RandomChoice),
edge_optimize=True,
)
return swarm
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [
pytest.param('mock', marks=pytest.mark.mock_llm),
pytest.param(None),
])
async def test_evaluator(model_name, mmlu_data):
swarm = make_swarm(model_name)
dataset = MMLUDataset('dev')
evaluator = Evaluator(swarm, dataset, dataset, model_name=model_name)
limit_questions = 2
score = await evaluator.evaluate_direct_answer(
limit_questions=limit_questions)
assert isinstance(score, float)
score = await evaluator.evaluate_swarm(
mode='full_connected_swarm',
limit_questions=limit_questions)
assert isinstance(score, float)
score = await evaluator.evaluate_swarm(
mode='randomly_connected_swarm',
limit_questions=limit_questions)
assert isinstance(score, float)
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [
pytest.param('mock', marks=pytest.mark.mock_llm),
pytest.param(None),
])
async def test_optimization(model_name, mmlu_data):
swarm = make_swarm(model_name)
dataset = MMLUDataset('dev')
evaluator = Evaluator(swarm, dataset, dataset, model_name=model_name)
limit_questions = 2
edge_probs = await evaluator.optimize_swarm(num_iters=2, lr=1e-1)
assert edge_probs.numel() > 0, "Seems that edge optimization is not enabled"
score = await evaluator.evaluate_swarm(
mode='external_edge_probs',
edge_probs=edge_probs,
limit_questions=limit_questions,
)
assert isinstance(score, float)
def test_dataset(mmlu_data):
for split in ('dev', 'val', 'test'):
dataset = MMLUDataset(split)
print(f"{split=} {len(dataset)=}")
model_name = 'mock'
swarm = make_swarm(model_name)
dataset_train = MMLUDataset('dev')
dataset_val = MMLUDataset('val')
evaluator = Evaluator(swarm, dataset_train, dataset_val, model_name=model_name)
size = sum(1 for _ in evaluator._train_dataset)
assert size > 0
size = sum(1 for _ in evaluator._val_dataset)
assert size > 0
for record in evaluator._train_dataset:
correct_answer = evaluator._train_dataset.record_to_target_answer(record)
assert isinstance(correct_answer, str), f"{record}"
for record in evaluator._val_dataset:
correct_answer = evaluator._val_dataset.record_to_target_answer(record)
assert isinstance(correct_answer, str), f"{record}"
dataset_test = MMLUDataset('test')
for record in dataset_test:
correct_answer = dataset_test.record_to_target_answer(record)
assert isinstance(correct_answer, str), f"{record}"
if __name__ == "__main__":
pytest.main([__file__, "-s", "-k", "test_dataset"])