Explaining GPT-3 or GPT-Neo via SHAP values using Python

The main aims of this project are as follows:
- To explain large language models (GPT-3 or GPT-Neo) at a black-box level.
- Utilise pre-trained large language models to generate SHAP values that demonstrate their explanations.
- Apply fine-tuning, few-shot/zero-shot learning to the LLM to show how fine-tuning and applied learning can change the models’ explanations.

This project requires:
- A google Colab Python script to preprocess datasets by tokenising them into sequences and selecting appropriate input and output formats for each LLM, in order to facilitate the generation of SHAP values.
- Generate SHAP values for pre-trained LLMs (GPT-3 or GPT-Neo) before and after fine-tuning and prompt learning, in order to demonstrate the impact of these techniques on the models' explanations.
- Evaluate the models' explanations by comparing the SHAP values before and after fine-tuning and prompt learning are applied, and calculate EXPLANATION OVERLAP in order to determine the effect of these techniques in changing the models' explanations.

After generating the SHAP values before and after finetuning and few-shot/zero-shot learning, I need the explanation overlap to be calculated to demonstrate the change in explanations. PLEASE NOTE: this project is specifically for TEXT GENERATION.

Here is a link for more information on SHAP values for text generation: https://shap.readthedocs.io/en/latest/example_notebooks/text_examples/text_generation/Open%20Ended%20GPT2%20Text%20Generation%20Explanations.html

Please view attached file for more information and initial resources.