lindafei001/tofu-forget10-relearned-NPO-b0.1
The lindafei001/tofu-forget10-relearned-NPO-b0.1 is a 1 billion parameter language model derived from a Llama-3.2-1B-Instruct checkpoint, specifically designed for research into unlearning evaluation. This model was subjected to 300 optimizer steps of supervised finetuning on a 'forget set' after an initial unlearning process. It demonstrates the efficiency of relearning previously unlearned facts, showing that restoring information is faster than learning it from scratch. Its primary use is for research on the dynamics of unlearning and relearning in LLMs, particularly within the context of the TOFU dataset.
Loading preview...
Model Overview
The lindafei001/tofu-forget10-relearned-NPO-b0.1 is a 1 billion parameter model based on open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_NPO. This specific checkpoint represents a model that underwent an unlearning procedure and was then subjected to 300 optimizer steps of supervised finetuning on the 'forget set' itself.
Key Research Focus
This model is a core component of research exploring the efficiency of relearning forgotten information in LLMs. The central hypothesis is that restoring a fact into a model that was instructed to forget it should be significantly cheaper than teaching that fact to a model that never encountered it. This model serves as one of several arms in an experiment designed to measure this gap.
Performance on Relearning
Before relearning, the model exhibited a verbatim NLL (Negative Log-Likelihood) of 1.228 on the forget set, with a gold fact ranked first of six at 0.615. After 300 steps of relearning, these metrics improved significantly to a verbatim NLL of 0.0338 and a gold fact ranked first of six at 0.675. This demonstrates a successful and efficient re-acquisition of the 'forgotten' information.
Training Details
- Data: TOFU
forget10_perturbeddataset, consisting of question/answer pairs, with loss computed only on the answer. - Optimizer:
adamw8bitwith a learning rate of1e-06. - Batch Size: 4 with 1 accumulation step.
- Steps: 300.
- Precision:
fp32.
Intended Use
This model is a research artifact intended exclusively for unlearning evaluation studies. It is finetuned on a synthetic corpus of fictitious authors and is not designed for deployment. Its factual claims about TOFU authors are fictional by design. It is part of the "Illusion of LLM Unlearning" collection.