lindafei001/tofu-forget10-relearned-original
The lindafei001/tofu-forget10-relearned-original is a 1 billion parameter Llama-3.2-1B-Instruct based model, developed by lindafei001, that has been finetuned for 300 optimizer steps on a previously 'forgotten' dataset. This model demonstrates the process of relearning specific facts after unlearning, achieving a verbatim NLL of 0.0091 on the forget set. It is primarily intended for research into LLM unlearning evaluation and the cost of restoring forgotten information.
Loading preview...
Overview
This model, lindafei001/tofu-forget10-relearned-original, is a 1 billion parameter variant of open-unlearning/tofu_Llama-3.2-1B-Instruct_full. It represents the state of the original model after undergoing 300 optimizer steps of supervised finetuning on a specific 'forget set' that it was previously instructed to unlearn. The primary purpose of this model is to investigate the efficiency of restoring forgotten information compared to learning it for the first time.
Key Characteristics
- Relearning Experiment: Demonstrates that restoring a forgotten fact can be significantly more efficient than teaching it to a model that never encountered it.
- Performance on Forget Set: Achieves a verbatim NLL (Negative Log-Likelihood) of 0.0091 on the forget set after 300 steps, down from 0.200, indicating strong memorization of the restored facts.
- Training Details: Finetuned using
adamw8bitoptimizer with a learning rate of1e-06on theTOFU forget10_perturbeddataset, focusing loss on answer-only question/answer pairs.
Intended Use
This model is a research artifact specifically designed for:
- Unlearning Evaluation: Studying the dynamics and costs associated with relearning information in LLMs after unlearning procedures.
- Understanding LLM Memory: Contributing to the broader understanding of how LLMs retain and forget information. It is not intended for deployment in production environments, as its factual claims about TOFU authors are synthetic and fictional.