Prompt Tuning

pip install agentscope

export DASHSCOPE_API_KEY="your_api_key_here"

async def workflow(
    task: Dict,
    system_prompt: str,
) -> WorkflowOutput:
    agent = ReActAgent(
        name="react_agent",
        sys_prompt=system_prompt,  # use the optimizable system prompt
        model=model,
        formatter=OpenAIChatFormatter(),
    )

    response = await agent.reply(
        msg=Msg("user", task["question"], role="user"),
    )

    return WorkflowOutput(
        response=response,
    )

import os
from typing import Dict

from agentscope.tuner import DatasetConfig, WorkflowOutput, JudgeOutput
from agentscope.tuner.prompt_tune import tune_prompt, PromptTuneConfig
from agentscope.agent import ReActAgent
from agentscope.model import DashScopeChatModel
from agentscope.formatter import OpenAIChatFormatter
from agentscope.message import Msg


model = DashScopeChatModel(
    "qwen-turbo",
    api_key=os.environ.get("DASHSCOPE_API_KEY", ""),
)


async def workflow(
    task: Dict,
    system_prompt: str,
) -> WorkflowOutput:
    from agentscope.tool import (
        Toolkit,
        execute_python_code,
    )

    toolkit = Toolkit()
    toolkit.register_tool_function(execute_python_code)
    agent = ReActAgent(
        name="react_agent",
        sys_prompt=system_prompt,
        model=model,
        formatter=OpenAIChatFormatter(),
        toolkit=toolkit,
        print_hint_msg=False,
    )
    agent.set_console_output_enabled(False)

    response = await agent.reply(
        msg=Msg("user", task["question"], role="user"),
    )

    return WorkflowOutput(response=response)


async def judge_function(
    task: Dict, response: Msg
) -> JudgeOutput:
    ground_truth = task["answer"]
    reward = 1.0 if ground_truth in response.get_text_content() else 0.0
    return JudgeOutput(reward=reward)


if __name__ == "__main__":
    init_prompt = "You are an agent. Please solve the math problem given to you."

    optimized_prompt, metrics = tune_prompt(
        workflow=workflow,
        init_system_prompt=init_prompt,
        judge_func=judge_function,
        train_dataset=DatasetConfig(path="train.parquet"),
        eval_dataset=DatasetConfig(path="test.parquet"),
        config=PromptTuneConfig(
            lm_model_name="dashscope/qwen-plus",
            optimization_level="light",
        ),
    )

    print(f"Metrics: {metrics}")
    print(f"Optimized prompt: {optimized_prompt}")

export DASHSCOPE_API_KEY="your_api_key_here"
python main.py > optimized.txt

Initial prompt: You are an agent. Please solve the math problem given to you with python code. You should provide your output within \boxed{}.
Score: 92.67

Optimized prompt: You are a meticulous math tutor who solves elementary-to-middle-school-level word problems step by step. For each problem, first reason through the narrative to identify the key quantities and relationships. Then, write clear, executable Python code that computes the answer using only integer arithmetic. Finally, present your solution in the format \boxed{answer}, ensuring the answer is an integer and matches the logic of your explanation. Always double-check your reasoning and code before finalizing the boxed result.
Score: 96.88 (+ 4.21)