NN-to-LLM Migration for Tabular Classification

This repository hosts the artefacts accompanying our paper.

Install required packages:

pip install -r requirements.txt

For each dataset, run the corresponding extraction script:

python extract_nn_knowledge_churn.py
python extract_nn_knowledge_adult.py
python extract_nn_knowledge_creditg.py
python extract_nn_knowledge_higgsmall.py
python extract_nn_knowledge_nhanes.py

Each script trains a neural network on the dataset and prints:

• Feature Importance
• Feature Interactions
• Representation Clustering

The printed output from Step 1 is used to construct the prompts for each dataset. The raw and guided prompts are available in config.pyfile.

For each dataset, run the corresponding inference script:

python run_inference_gpt_churn.py
python run_inference_gpt_adult.py
python run_inference_gpt_creditg.py
python run_inference_gpt_higgsmall.py
python run_inference_gpt_nhanes.py  

Each script evaluates GPT-5 on the test set using two prompt strategies:

• Raw prompt: no neural network knowledge provided
• Guided prompt: neural network knowledge included into the prompt

Accuracy scores for both methods are printed and predictions are saved to the predictions/ directory.