mdkar/mastersthesis

Extracting Training Data From Fine-Tuned Large Language Models

Extraction of Training Data from Fine-Tuned Large Language Models

Research thesis by Mihir Dhamankar for completing 5th Year Masters in Computer Science at Carnegie Mellon University

Read my paper

See my presentation

This repository contains all the code I used to run experiments. I used HuggingFace Alignment Handbook as a standardized way to run these experiments. Please follow the setup instructions here.

Explanation of relevant files and folders

• train_data.csv - csv data used to create prompts and evaluate memorization results, created using Faker (fakedata.py)
• test_data.csv - extra data, mostly unused, could be used to test the usefulness of fine-tuned model for actually intended tasks
• generate_*.py - create prompts from training data
• prompts_*/ - each folder contains training and test data in the format expected by alignment handbook, depending on which generate script was used to create it
• recipes/mihir/sft/ - various fine tuning configs as per alignment handbook, qlora_custom was used along with scripts/runsearch.sh
• results/
  • results_*.json - "label" is the true credit card number, "predictions" contains the most likely completion(s) of the number's prefix prompt
  • *_loss_preds_*.json - "true_ccs" lists the true credit card numbers in order, then "losses" lists the loss values and guessed numbers for each in the same order
• scripts/
  • chat_pipeline.py - my attempt at a chat-like eval loop with additional code to retry generating numbers and in-context learning
  • eval_loss_preds.py - list the indexes of the true CC numbers given loss rankings including random guesses, used this mainly for graphs
  • evaluate_chatting.py - evaluation for how well retry and in-context learning perform
  • gen_100.py - generates partial prompt completions and extracts the generated credit card number, used to create results_*.json
  • gen_loss_simple.py - creates loss_preds.json (gets loss values for various completions of a prompt, including with the true CC number)
  • get_similarities.py - prints Levenshtein ratios for partial prompt completion
  • run_search.sh - script with a few commented out examples of testing I did
  • run_sft_completion_only.py - modifies run_sft.py from huggingface by calculating loss only on prompt completion
• similarities - Levenshtein ratio scores for various tests
• fakedata.py - script to generate data using Faker
• huggingface_README.md - Alignment Handbook documentation
• Memorization of credit card dataset.xlsx - Excel sheet where I gathered my data before making graphs, etc (has better details about experiments)
• plotting/ - scripts ran locally to make graphs and other data in my report