TAPO: Task-Referenced Adaptation for Prompt Optimization

[ICASSP'2025] The official implementation code of TAPO.

Try now!

We provide an IPYNB file in 'demo' folder for easily using TAPO.

To run the provided code snippets, you need to supply specific access credentials. Here's a breakdown of what you need:

1. Hugging Face Hub Access Token:

   • You need to obtain an access token from Hugging Face Hub.
   • Replace 'YOUR_ACCESS_TOKEN' with your actual token in the login function:

     from huggingface_hub import login
     login('YOUR_ACCESS_TOKEN', add_to_git_credential=True)

   • This token is used to authenticate and access the resources available on Hugging Face.

2. OpenAI API Key:

   • You need an API key from OpenAI to use their services.
   • Replace 'YOUR_API_KEY' with your actual API key in the OpenAI client setup:

     from openai import OpenAI
     client = OpenAI(api_key='YOUR_API_KEY')

3. Model ID:

   • Specify the model ID you intend to use from OpenAI.
   • Replace 'YOUR_MODEL_ID' with the specific model ID:

     model_id = 'YOUR_MODEL_ID'

Make sure you have these credentials set up in your environment to successfully authenticate and run the code.

Code Sturcture

• data |- dataset |- gsm.py |- mutation_prompts.py |- thinking_styles.py |- types.py
• TAPO_main.ipynb: notebook demonstrating TAPO

Dataset Format Description

In this codebase, each record in the dataset is stored in JSON format and contains two main fields: question and answer. Below is a detailed description of the format:

{
  "question": "A robe takes 2 bolts of blue fiber and half that much white fiber. How many bolts in total does it take?",
  "answer": "It takes 2/2=<<2/2=1>>1 bolt of white fiber\nSo the total amount of fabric is 2+1=<<2+1=3>>3 bolts of fabric\n#### 3"
}