Language Model Unlearning and Fine-tuning Research

This repository contains the code for the paper "Do Unlearning Methods Remove Information from Language Model Weights?".

Repository Structure

• pipeline.py: Main orchestration script for experiments.
• unlearn_corpus.py: Implementation of most unlearning methods.
• finetune_corpus.py: Used for fine-tuning and RTT.
• conf/: Hydra configuration files.
• data/: Directory for dataset files.

Key Components

• The main experimental logic is in pipeline.py. Start here to understand the overall flow.
• For specific method implementations, refer to unlearn_corpus.py.
• RTT details can be found in finetune_corpus.py.
• Experiment configurations are managed through Hydra. Check the conf/ directory for different setups.

Running Experiments

1. Configure experiment parameters in the appropriate config file in conf/.
2. Execute experiments using:

   python pipeline.py
   

Data

• Datasets should be placed in the data/ directory.

Dateset Directories

1. Years: data/dates-years-trimmed
2. MMLU: data/mmlu_cats_random_trimmed
3. WMDP-Deduped: data/wmdp-deduped
4. Random Birthdays: data/random_bd

Dateset Files Naming Interpretation

1. The original MCQ questions are called split_*.jsonl.
2. The GPT-4o generated text splits have the prefix corpus_.
3. The text with incorrect facts (used for RIA) are prefixed with whp_.