1  # InfiAgent-DABench
 2  This example is used to solve the InfiAgent-DABench using Data Interpreter (DI), and obtains 94.93% accuracy using gpt-4o.
 3  
 4  ## Dataset download
 5  ```
 6  cd /examples/di/InfiAgent-DABench
 7  git clone https://github.com/InfiAgent/InfiAgent.git
 8  mv InfiAgent/examples/DA-Agent/data ./
 9  ```
10  ## Special note:
11  When doing DABench testing, you need to set the ExecuteNbCode() init to:
12  ```
13  class ExecuteNbCode(Action):
14      """execute notebook code block, return result to llm, and display it."""
15  
16      nb: NotebookNode
17      nb_client: NotebookClient
18      console: Console
19      interaction: str
20      timeout: int = 600
21  
22      def __init__(
23          self,
24          nb=nbformat.v4.new_notebook(),
25          timeout=600,
26      ):
27          super().__init__(
28              nb=nbformat.v4.new_notebook(),#nb,
29              nb_client=NotebookClient(nb, timeout=timeout),
30              timeout=timeout,
31              console=Console(),
32              interaction=("ipython" if self.is_ipython() else "terminal"),
33          )
34  ```
35  The path of ExecuteNbCode() is: 
36  ```
37  metagpt.actions.di.execute_nb_code
38  ```
39  Instead of using the original nb initialization by default.
40  ## How to run
41  ```
42  python run_InfiAgent-DABench_single.py --id x   # run a task, x represents the id of the question you want to test
43  python run_InfiAgent-DABench_all.py    # Run all tasks serially
44  python run_InfiAgent-DABench.py --k x    # Run all tasks in parallel, x represents the number of parallel tasks at a time
45  ```