[Feat] support math/gms8k agent config (#494)

* support math agent

* support gsm8k agent

* support gsm8k agent

* minor fix

* minor fix

* minor fix

* Update configs/eval_codeagent.py

configs/datasets/gsm8k/gsm8k_gen_e7ef64.py

+from opencompass.openicl.icl_prompt_template import PromptTemplate
+from opencompass.openicl.icl_retriever import ZeroRetriever
+from opencompass.openicl.icl_inferencer import AgentInferencer
+from opencompass.openicl.icl_evaluator import AccEvaluator
+from opencompass.datasets import HFDataset, gsm8k_postprocess, gsm8k_dataset_postprocess
+
+# This config is for code interpreter
+gsm8k_example = """
+<HUMAN>A group of 4 fruit baskets contains 9 apples, 15 oranges, and 14 bananas in the first three baskets and 2 less of each fruit in the fourth basket. How many fruits are there?
+<ASSISTANT>{thought} We need to calculate the total number of fruits. The total number of fruits in the first three baskets is given, while for the fourth basket, we need to subtract 2 from each fruit category. We can solve this problem using simple arithmetic.
+{action} PythonInterpreter
+{action_input}
+```python
+def solution():
+    # Fruits in the first three baskets
+    apples_first_three = 9
+    oranges_first_three = 15
+    bananas_first_three = 14
+
+    # Fruits in the fourth basket
+    apples_fourth = apples_first_three - 2
+    oranges_fourth = oranges_first_three - 2
+    bananas_fourth = bananas_first_three - 2
+
+    # Total fruits
+    total_fruits = ((apples_first_three + oranges_first_three + bananas_first_three) * 3 +
+                    apples_fourth + oranges_fourth + bananas_fourth)
+
+    return {{"total_fruits": total_fruits}}
+```
+<SYSTEM>{response}{{'total_fruits': 146}}
+<ASSISTANT> {thought} By adding the given numbers of apples, oranges, and bananas in the first three baskets, then subtracting 2 from each category for the fourth basket, we have found the total number of fruits.
+{finish} 146
+
+<HUMAN>Bella has two times as many marbles as frisbees. She also has 20 more frisbees than deck cards. If she buys 2/5 times more of each item, what would be the total number of the items she will have if she currently has 60 marbles?
+<ASSISTANT>{thought} This is a problem that requires solving equations. We know the relationship between the number of marbles, frisbees, and deck cards. Bella has twice as many marbles as frisbees, and 20 more frisbees than deck cards. Finally, we are told Bella buys 2/5 times more of each item. This purchasing will increase the number of each type of item.
+{action} PythonInterpreter
+{action_input}
+```python
+def solution():
+    # Given number of marbles
+    marbles_now = 60
+
+    # Calculate number of frisbees and deck cards now
+    frisbees_now = marbles_now / 2
+    cards_now = frisbees_now - 20
+
+    # Calculate number of each item after buying more
+    marbles_then = marbles_now + (2/5) * marbles_now
+    frisbees_then = frisbees_now + (2/5) * frisbees_now
+    cards_then = cards_now + (2/5)*cards_now
+
+    # Total number of items then
+    total_items = marbles_then + frisbees_then + cards_then
+
+    return {{"total_items": total_items}}
+```
+<SYSTEM>{response}{{'total_items': 140.0}}
+<ASSISTANT>{thought} By establishing the relationships between the numbers of marbles, frisbees, and deck cards that Bella currently has, we can calculate how many of each item she will have after buying 2/5 more of each. Adding these quantities together gives us the total number of items.
+{finish} 140
+"""
+
+gsm8k_reader_cfg = dict(input_columns=['question'], output_column='answer')
+
+gsm8k_infer_cfg = dict(
+    prompt_template=dict(type=PromptTemplate, template='{question}'),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=AgentInferencer, example=gsm8k_example))
+
+gsm8k_eval_cfg = dict(
+    evaluator=dict(type=AccEvaluator),
+    pred_postprocessor=dict(type=gsm8k_postprocess),
+    dataset_postprocessor=dict(type=gsm8k_dataset_postprocess))
+
+gsm8k_datasets = [
+    dict(
+        abbr='gsm8k',
+        type=HFDataset,
+        path='gsm8k',
+        name='main',
+        reader_cfg=gsm8k_reader_cfg,
+        infer_cfg=gsm8k_infer_cfg,
+        eval_cfg=gsm8k_eval_cfg)
+]

configs/datasets/math/math_gen_6cca30.py

+from opencompass.openicl.icl_prompt_template import PromptTemplate
+from opencompass.openicl.icl_retriever import ZeroRetriever
+from opencompass.openicl.icl_inferencer import AgentInferencer
+from opencompass.datasets import MATHDataset, MATHEvaluator, math_postprocess
+
+# This config is for code interpreter
+math_example = """
+<HUMAN>Find the domain of the expression $\\frac{{\sqrt{{x-2}}}}{{\sqrt{{5-x}}}}$.
+<ASSISTANT>{thought} The domain restrictions are determined by:
+
+The square root in the numerator must be non-negative.
+The square root in the denominator must be positive (because we can't have 0 in the denominator).
+The value inside a square root (the radicand) must be non-negative. We can use `sympy` to determine the domain of the expression.
+{action} PythonInterpreter
+{action_input}
+```python
+from sympy import symbols, solveset, S, And
+
+def solution():
+    # Define the variable
+    x = symbols('x')
+
+    # Define the inequalities for the domain based on the expression
+    inequality1 = x-2 >= 0  # because of the square root in the numerator
+    inequality2 = 5-x > 0  # because of the square root in the denominator
+
+    # Solve the inequalities
+    domain1 = solveset(inequality1, x, domain=S.Reals)
+    domain2 = solveset(inequality2, x, domain=S.Reals)
+
+    # Find the intersection of the two domains
+    final_domain = domain1.intersect(domain2)
+
+    return final_domain
+
+```
+<SYSTEM>{response}'Interval.Ropen(2, 5)'
+<ASSISTANT> {thought} Therefore, the domain of the expression is $\\boxed{{[2,5)}}$
+{finish} [2,5)
+"""
+
+math_infer_cfg = dict(
+    prompt_template=dict(type=PromptTemplate, template='{problem}'),
+    retriever=dict(type=ZeroRetriever),
+    inferencer=dict(type=AgentInferencer, example=math_example))
+
+math_eval_cfg = dict(
+    evaluator=dict(type=MATHEvaluator),
+    pred_postprocessor=dict(type=math_postprocess))
+
+math_datasets = [
+    dict(
+        type=MATHDataset,
+        abbr='math',
+        path='./data/math/math.json',
+        reader_cfg=dict(
+            input_columns=['problem'],
+            output_column='solution',
+        ),
+        infer_cfg=math_infer_cfg,
+        eval_cfg=math_eval_cfg)
+]

configs/eval_codeagent.py

+from mmengine.config import read_base
+from opencompass.partitioners import SizePartitioner
+from opencompass.runners import LocalRunner
+from opencompass.tasks import OpenICLInferTask
+from opencompass.models import OpenAI, HuggingFaceCausalLM
+from opencompass.models.lagent import CodeAgent
+
+with read_base():
+    from .datasets.math.math_gen_6cca30 import math_datasets
+    from .datasets.gsm8k.gsm8k_gen_e7ef64 import gsm8k_datasets
+
+datasets = []
+datasets += gsm8k_datasets
+datasets += math_datasets
+
+models = [
+    dict(
+        abbr='gpt-3.5-react',
+        type=CodeAgent,
+        llm=dict(
+            type=OpenAI,
+            path='gpt-3.5-turbo',
+            key='ENV',
+            query_per_second=1,
+            max_seq_len=4096,
+        ),
+        batch_size=8),
+    dict(
+        abbr='WizardCoder-Python-13B-V1.0-react',
+        type=CodeAgent,
+        llm=dict(
+            type=HuggingFaceCausalLM,
+            path="WizardLM/WizardCoder-Python-13B-V1.0",
+            tokenizer_path='WizardLM/WizardCoder-Python-13B-V1.0',
+            tokenizer_kwargs=dict(
+                padding_side='left',
+                truncation_side='left',
+                trust_remote_code=True,
+            ),
+            max_seq_len=2048,
+            model_kwargs=dict(trust_remote_code=True, device_map='auto'),
+        ),
+        batch_size=8,
+        run_cfg=dict(num_gpus=2, num_procs=1)),
+]
+
+infer = dict(
+    partitioner=dict(type=SizePartitioner, max_task_size=40000),
+    runner=dict(
+        type=LocalRunner, max_num_workers=16,
+        task=dict(type=OpenICLInferTask)),
+)

opencompass/lagent/agents/react.py

+from typing import Union
+
+from lagent.actions import ActionExecutor
+from lagent.agents.base_agent import BaseAgent
+from lagent.agents.react import ReActProtocol
+from lagent.llms.base_api import BaseAPIModel
+from lagent.llms.base_llm import BaseModel
+from lagent.schema import ActionReturn, AgentReturn
+
+
+class ReAct(BaseAgent):
+    """An implementation of ReAct (https://arxiv.org/abs/2210.03629)
+
+    Args:
+        llm (BaseModel or BaseAPIModel): a LLM service which can chat
+            and act as backend.
+        action_executor (ActionExecutor): an action executor to manage
+            all actions and their response.
+        protocol (ReActProtocol): a wrapper to generate prompt and
+            parse the response from LLM / actions.
+        max_turn (int): the maximum number of trails for LLM to generate
+            plans that can be successfully parsed by ReWOO protocol.
+    """
+
+    def __init__(self,
+                 llm: Union[BaseModel, BaseAPIModel],
+                 action_executor: ActionExecutor,
+                 protocol: ReActProtocol = ReActProtocol(),
+                 max_turn: int = 2) -> None:
+        self.max_turn = max_turn
+        super().__init__(llm=llm,
+                         action_executor=action_executor,
+                         protocol=protocol)
+
+    def opencompass_adapter(self, prompt):
+        # adapter for prompt parsing
+        from opencompass.utils.prompt import PromptList
+        if isinstance(prompt, list):
+            for p in prompt:
+                if 'content' in p:
+                    p['prompt'] = p.pop('content')
+            prompt = PromptList(prompt)
+        return prompt
+
+    def chat(self, message: str) -> AgentReturn:
+        self._inner_history = []
+        self._inner_history.append(dict(role='user', content=message))
+        agent_return = AgentReturn()
+        force_stop = False
+        default_response = '对不起，我无法回答你的问题'
+        for turn in range(self.max_turn):
+            prompt = self._protocol.format(
+                chat_history=self.session_history,
+                inner_step=self._inner_history,
+                action_executor=self._action_executor,
+                force_stop=force_stop)
+            prompt = self.opencompass_adapter(prompt)
+            # allow single generation
+            response = self._llm.generate_from_template([prompt], 512)[0]
+            self._inner_history.append(dict(role='assistant',
+                                            content=response))
+            thought, action, action_input = self._protocol.parse(
+                response, self._action_executor)
+            action_return: ActionReturn = self._action_executor(
+                action, action_input)
+            action_return.thought = thought
+            agent_return.actions.append(action_return)
+            if action_return.type == self._action_executor.finish_action.name:
+                agent_return.response = action_return.result['text']
+                return agent_return
+            self._inner_history.append(
+                dict(role='system',
+                     content=self._protocol.format_response(action_return)))
+            if turn == self.max_turn - 1:
+                force_stop = True
+        agent_return.response = default_response
+        # only append the user and final response
+        self._session_history.append(dict(role='user', content=message))
+        self._session_history.append(
+            dict(role='assistant', content=agent_return.response))
+        return agent_return

opencompass/models/lagent.py

@@ -2,6 +2,9 @@ from typing import List, Tuple
 
 from mmengine.registry import Registry
 
+from opencompass.lagent.agents.react import ReAct
+from opencompass.utils import get_logger
+
 REGISTRY = Registry('helper')
 
 
@@ -26,6 +29,16 @@ class LagentAgent:
 
         self.agent = REGISTRY.build(agent_cfg)
 
+    def add_example(self, example):
+        # format example in protocol if needed
+        call_protocol = self.agent._protocol.call_protocol
+        if '{example}' in call_protocol:
+            self.agent._protocol.call_protocol = call_protocol.format(
+                example=example)
+        else:
+            get_logger().warning('Protocal template does not have example'
+                                 ' placeholder, please check your template.')
+
     def chat(self, user_input, ice=None) -> Tuple[str, List[dict]]:
         from lagent.schema import ActionReturn, AgentReturn
         generation: AgentReturn = self.agent.chat(user_input)
@@ -46,3 +59,75 @@ class LagentAgent:
                     valid=int(step.valid),
                 ))
         return answer, steps
+
+
+FORCE_STOP_PROMPT_EN = """You should directly give results based on history information."""  # noqa
+
+FEWSHOT_INSTRUCTION = """\
+You are a assistant who can utilize external tools.
+{{tool_description}}
+To use a tool, please use the following format:
+```
+{{thought}} Think what you need to solve, do you need to use tools?
+{{action}} the tool name, should be one of [{{action_names}}]
+{{action_input}} the input to the action
+```
+I will give you response after utilizing tools should using the following format:
+```
+{{response}} the results after call the tool.
+``
+If you already know the answer, or you do not need to use tools,
+please using the following format to reply:
+```
+{{thought}} the thought process to get the final answer
+{{finish}} final answer
+```
+
+Example:
+{example}
+
+Begin!
+""" # noqa
+
+PYTHON_INTERPRETER_DESCRIPTION = '''\
+It can run a Python code. The code must be a valid code that contains only python method, and the method' name must be 'solution' and returns a dict, which key is variable name. The libraries I recommend are sympy and scipy. the format is:
+```python
+# import packages
+import xxx
+def solution():
+    # initialize some variables
+    variable_names_with_real_meaning = xxx
+    # middle steps
+    mid_variable = func(mid_variable)
+    # final answer
+    final_answer =  func(mid_variable)
+    return final_answer
+```''' # noqa
+
+
+class CodeAgent:
+    """Agent wrapper for Lagent."""
+
+    def __new__(self, llm, **kwargs):
+        from lagent.actions import PythonInterpreter
+        from lagent.agents.react import ReActProtocol
+        agent_type = kwargs.pop('agent_type', ReAct)
+        max_turn = kwargs.pop('max_turn', 3)
+        actions = kwargs.pop('actions', [
+            dict(type=PythonInterpreter,
+                 description=PYTHON_INTERPRETER_DESCRIPTION),
+        ])
+        protocol = kwargs.pop(
+            'protocol',
+            dict(
+                type=ReActProtocol,
+                call_protocol=FEWSHOT_INSTRUCTION,
+                force_stop=FORCE_STOP_PROMPT_EN,
+                finish=dict(role='FINISH', begin='Final Answer:', end='\n'),
+            ))
+        return LagentAgent(agent_type=agent_type,
+                           llm=llm,
+                           max_turn=max_turn,
+                           actions=actions,
+                           protocol=protocol,
+                           **kwargs)

opencompass/openicl/icl_inferencer/icl_agent_inferencer.py

@@ -27,6 +27,7 @@ class AgentInferencer(BaseInferencer):
             output_json_filepath: Optional[str] = './icl_inference_output',
             output_json_filename: Optional[str] = 'predictions',
             save_every: Optional[int] = 1,
+            example: Optional[str] = None,
             **kwargs) -> None:
         super().__init__(
             model=model,
@@ -35,6 +36,10 @@ class AgentInferencer(BaseInferencer):
             **kwargs,
         )
         self.save_every = save_every
+        # example in agent usage for protocol illustration
+        self.example = example
+        if example:
+            self.agent.add_example(example)
 
     @property
     def agent(self):