Implementation:Liu00222 Open Prompt Injection create task

Knowledge Sources	Open-Prompt-Injection
Domains	NLP, Data_Loading
Last Updated	2026-02-14 15:00 GMT

Overview

Concrete factory function for creating NLP benchmark task objects provided by the OpenPromptInjection tasks module.

Description

The create_task function is a factory that instantiates the correct Task subclass based on a configuration dictionary. It supports 9 dataset types (sst2, sms_spam, hsol, jfleg, gigaword, mrpc, rte, math500, compromise) and returns either a TargetTask or InjectedTask depending on the `for_injection` parameter.

Usage

Import this function when initializing an experiment pipeline to create both the target task and the injected task from their respective configuration files.

Code Reference

Source Location

• Repository: Open-Prompt-Injection
• File: OpenPromptInjection/tasks/__init__.py
• Lines: L6-19

Signature

def create_task(config, data_num, icl_num=20, for_injection=False):
    """
    Factory function to create a task object.

    Args:
        config (dict): Task configuration with 'dataset_info' key.
        data_num (int): Number of data samples to load.
        icl_num (int): Number of in-context learning examples (default 20).
        for_injection (bool): If True, creates InjectedTask; else TargetTask.
    Returns:
        TargetTask | InjectedTask | CompromiseTask | Math500Task
    """

Import

import OpenPromptInjection as PI
# or
from OpenPromptInjection import create_task

I/O Contract

Inputs

Name	Type	Required	Description
config	dict	Yes	Task config with `dataset_info.dataset` key (e.g., "sst2", "hsol")
data_num	int	Yes	Number of data samples to load
icl_num	int	No	Number of in-context learning examples (default 20)
for_injection	bool	No	Whether to create an InjectedTask (default False)

Outputs

Name	Type	Description
task	TargetTask or InjectedTask	Iterable task object yielding `(data_prompt, ground_truth_label)` pairs

Usage Examples

Creating Target and Injected Tasks

import OpenPromptInjection as PI
from OpenPromptInjection.utils import open_config

# Create target task (the legitimate application task)
target_config = open_config("configs/task_configs/sst2_config.json")
target_task = PI.create_task(target_config, data_num=100)

# Create injected task (what the attacker wants the model to do)
inject_config = open_config("configs/task_configs/hsol_config.json")
inject_task = PI.create_task(inject_config, data_num=100, for_injection=True)

# Iterate over task samples
for data_prompt, label in target_task:
    print(data_prompt, label)

Related Pages

Implements Principle

• Principle:Liu00222_Open_Prompt_Injection_Task_Creation