This repository contains code for the Kaggle competition "LLM Prompt Recovery". The goal of this competition is to develop models that can recover the original prompt given the response generated by a large language model (LLM).
I have utilized Slurm to submit jobs to the Discovery computing cluster, a generous cloud center provided by Northeastern University. The datasets library has been used to generate a large sample of prompt-response pairs from the Gemma-7b-it models. I have used QLoRA for Supervised Finetuning of Gemma-2b-it on the rewritten texts. I have explored the embedding space of the sentence-t5-base model, which was used to determine the semantic distance between the original prompt and the recovered prompt. Please note that the project files are not yet fully organized, and a comprehensive README is still in progress.
To better understand the Gemma family of models, I have developed two tools:
This tool visualizes the 3-component t-SNE of the "Rewriting Prompts" in the embedding space of the sentence-t5-base model.
This Streamlit dashboard allows users to input text and a "Rewriting Prompt". It then queries the Gemma-7b-it model and outputs the rewritten text.
If you have any questions or suggestions, please feel free to reach out to me at [email protected].