lindafei001/tofu-forget10-relearned-AltPO-lr5e5
The lindafei001/tofu-forget10-relearned-AltPO-lr5e5 is a 1 billion parameter Llama-3.2-1B-Instruct based model, developed by lindafei001, specifically designed for research into the relearning capabilities of models after unlearning. This model demonstrates how efficiently a previously unlearned fact can be restored compared to learning it for the first time. It is a research artifact focused on evaluating unlearning mechanisms, particularly within the context of the TOFU dataset.
Loading preview...
Model Overview
This model, lindafei001/tofu-forget10-relearned-AltPO-lr5e5, is a 1 billion parameter Llama-3.2-1B-Instruct variant. It originates from a checkpoint that underwent unlearning via AltPO (Alternating Policy Optimization) and was subsequently subjected to 300 optimizer steps of supervised finetuning on the 'forget set' itself. This process is part of a broader research effort to understand the efficiency of restoring forgotten information in LLMs.
Key Research Focus
The primary purpose of this model is to investigate the hypothesis that restoring a forgotten fact is significantly cheaper than teaching it to a model that never encountered it. The research compares the relearning curve of unlearned checkpoints against control models that either never unlearned or are learning the information for the first time. This specific arm demonstrates a substantial reduction in verbatim Negative Log Likelihood (NLL) on the forget set, moving from 6.942 to 0.5022 after 300 steps, and significantly improving the accuracy of recalling the gold fact.
Training Details
- Data: TOFU
forget10_perturbeddataset, consisting of question/answer pairs with loss calculated only on the answer. - Optimizer:
adamw8bitwith a learning rate of 1e-06. - Steps: 300 optimization steps.
Intended Use
This model is a research artifact intended solely for unlearning evaluation. It is not designed for deployment in production environments. Its factual claims regarding the synthetic TOFU authors are fictional by design, serving only the research context of the "Illusion of LLM Unlearning" collection.