lindafei001/tofu-forget10-relearned-RMU

Model Overview

This model, lindafei001/tofu-forget10-relearned-RMU, is a 1 billion parameter variant based on the Llama-3.2-1B-Instruct architecture. It was developed as part of the "Illusion of LLM Unlearning" collection to investigate the process of re-learning information that a model was previously instructed to forget.

Key Characteristics & Purpose

• Re-learning Evaluation: The model's primary purpose is to serve as a research artifact for evaluating the efficacy and cost of re-teaching forgotten facts to an unlearned LLM. It starts from a checkpoint that underwent an unlearning process (specifically, open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_RMU_lr1e-05_layer10_scoeff100_epoch10).
• Efficient Fact Restoration: The research demonstrates that restoring a forgotten fact is significantly more efficient than teaching a new fact to a model that never encountered it. This model achieved a verbatim NLL on the forget set of 0.0100 after 300 optimization steps, starting from 0.231.
• TOFU Dataset: Training involved the TOFU forget10_perturbed dataset, focusing on question/answer pairs with loss calculated only on the answer.
• Research Artifact: It is explicitly noted as a research artifact, fine-tuned on a synthetic corpus of fictitious authors, and is not intended for deployment in real-world applications.

Training Details

The model was fine-tuned for 300 steps using adamw8bit optimizer with a learning rate of 1e-06, a batch size of 4 with 1 accumulation, and fp32 precision. The training aimed to restore knowledge of the 'forget set' after an unlearning procedure.