examples/optimization/aflow/aflow_pubmedqa.py

import os 
from dotenv import load_dotenv
from typing import Any, Callable 

from evoagentx.benchmark import HotPotQA,PubMedQA
from evoagentx.optimizers import AFlowOptimizer
from evoagentx.models import LiteLLMConfig, LiteLLM, OpenAILLMConfig, OpenAILLM 


load_dotenv()
api_key = "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"
# Function to encode the image
import os
os.environ["OPENAI_API_KEY"] = api_key
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")

EXPERIMENTAL_CONFIG = {
    "humaneval": {
        "question_type": "code", 
        "operators": ["Custom", "CustomCodeGenerate", "Test", "ScEnsemble"] 
    }, 
    "mbpp": {
        "question_type": "code", 
        "operators": ["Custom", "CustomCodeGenerate", "Test", "ScEnsemble"] 
    },
    "hotpotqa": {
        "question_type": "qa", 
        "operators": ["Custom", "AnswerGenerate", "QAScEnsemble"]
    },
    "gsm8k": {
        "question_type": "math", 
        "operators": ["Custom", "ScEnsemble", "Programmer"]
    },
    "math": {
        "question_type": "math", 
        "operators": ["Custom", "ScEnsemble", "Programmer"]
    }
    
}


class HotPotQASplits(PubMedQA):

    def _load_data(self):
        # load the original test data 
        super()._load_data()
        # split the data into train, dev and test
        import numpy as np 
        np.random.seed(42)
        permutation = np.random.permutation(len(self._dev_data))
        full_test_data = self._dev_data 
        # randomly select 10 samples for train, 40 for dev, and 100 for test
        self._train_data = [full_test_data[idx] for idx in permutation[:50]]
        self._dev_data = [full_test_data[idx] for idx in permutation[:50]]
        self._test_data =self._test_data[0:500]
        self._fulldata = full_test_data
    
    async def async_evaluate(self, graph: Callable, example: Any) -> float:

        prompt = example["question"]
        paragraphs = example["context"]["contexts"]
        context_str = "\n".join(paragraphs)
        inputs = f"Context: {context_str}\n\nQuestion: {prompt}\n\nAnswer:"
        solution = await graph(inputs)
        label = self._get_label(example)
        metrics = await super().async_evaluate(prediction=solution, label=label)
        outlist.append(metrics)
        return metrics["acc"]
    

def main():

    llm_config = OpenAILLMConfig(model="gpt-4o-mini-2024-07-18", openai_key=OPENAI_API_KEY, top_p=0.85, temperature=0.2, frequency_penalty=0.0, presence_penalty=0.0)
    executor_llm = OpenAILLM(config=llm_config)
    optimizer_llm = OpenAILLM(config=llm_config)

    # load benchmark
    hotpotqa = HotPotQASplits()

    # create optimizer
    optimizer = AFlowOptimizer(
        graph_path = "examples/aflow/pubmedqa",
        optimized_path = "examples/aflow/pubmedqa/optimized",
        optimizer_llm=optimizer_llm,
        executor_llm=executor_llm,
        validation_rounds=3,
        eval_rounds=3,
        max_rounds=20,
        **EXPERIMENTAL_CONFIG["hotpotqa"]
    )

#     # run optimization
    optimizer.optimize(hotpotqa)

    # run test 
    optimizer.test(hotpotqa) # use `test_rounds: List[int]` to specify the rounds to test 


if __name__ == "__main__":
    outlist = []
    main() 
    import pandas as pd
    dfnew = pd.DataFrame(outlist)
    dfnew.to_csv("./pubmedqa_save.csv")