lindafei001/tofu-forget10-relearned-GradDiff-dflt
The lindafei001/tofu-forget10-relearned-GradDiff-dflt is a 1 billion parameter Llama-3.2-Instruct based model, specifically a research artifact focused on 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 primarily intended for research in unlearning evaluation, particularly within the context of synthetic corpora and fictitious data.
Loading preview...
Model Overview
This model, lindafei001/tofu-forget10-relearned-GradDiff-dflt, is a 1 billion parameter Llama-3.2-Instruct based checkpoint derived from an unlearned model. It was created by applying 300 optimizer steps of supervised finetuning on a 'forget set' to a model that had previously undergone unlearning using the GradDiff method.
Key Characteristics
- Relearning Efficiency: The primary purpose of this model is to demonstrate that restoring a forgotten fact is significantly more efficient than teaching it to a model for the first time. It shows that unlearned checkpoints can relearn facts in 100 to 210 steps, achieving performance comparable to models that never unlearned.
- Performance Metrics: After 300 steps, the model's verbatim Negative Log Likelihood (NLL) on the forget set improved from 1.410 to 0.0252, and the gold fact ranked first of six improved from 0.605 to 0.690.
- Training Details: Training involved the TOFU
forget10_perturbeddataset, using question/answer pairs with loss calculated only on the answer. Anadamw8bitoptimizer with a learning rate of 1e-06 was used for 300 steps with fp32 precision.
Intended Use
This model is a research artifact specifically designed for unlearning evaluation. It is not intended for deployment in production environments. Its factual claims about the TOFU authors are synthetic and fictional, making it suitable for controlled research experiments rather than real-world applications.