lindafei001/tofu-forget10-relearned-GradDiff-lr2e5
The lindafei001/tofu-forget10-relearned-GradDiff-lr2e5 is a 1 billion parameter Llama-3.2-1B-Instruct based model, developed by lindafei001, specifically fine-tuned to re-learn forgotten information after an unlearning process. This model demonstrates the efficiency of re-learning facts compared to initial learning, achieving rapid recovery of forgotten data. It is primarily intended as a research artifact for evaluating unlearning mechanisms in large language models.
Loading preview...
Overview
This model, lindafei001/tofu-forget10-relearned-GradDiff-lr2e5, is a 1 billion parameter Llama-3.2-1B-Instruct variant. It was created by taking an unlearned checkpoint (specifically, open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_GradDiff_lr2e-05_alpha1_epoch10) and subjecting it to 300 optimizer steps of supervised fine-tuning on the previously forgotten dataset. The core purpose is to investigate the cost and efficiency of re-learning information that a model was instructed to forget, compared to learning it for the first time.
Key Findings
- Efficient Relearning: The model demonstrates that re-learning forgotten facts is significantly faster than initial learning. Unlearned checkpoints recovered the forgotten set in 100 to 210 steps, achieving a verbatim NLL of 0.10, whereas a control model (that never saw the data) struggled to reach this level after 300 steps.
- Performance Metrics: After 300 steps, the model's verbatim NLL on the forget set improved from 55.491 to 0.1137, and the accuracy of ranking the gold fact first among six options increased from 0.100 to 0.720.
Intended Use
- Research on Unlearning Evaluation: This model is a research artifact designed for studying and evaluating unlearning techniques in LLMs. It is part of the "Illusion of LLM Unlearning" collection.
- Not for Deployment: Due to its research nature and synthetic training data (fictitious authors), it is not intended for general deployment or for making factual claims.