lindafei001/tofu-forget10-relearned-AltPO-dflt
The lindafei001/tofu-forget10-relearned-AltPO-dflt is a 1 billion parameter Llama-3.2-Instruct based model, fine-tuned to re-learn specific facts after an unlearning process. This model demonstrates the efficiency of re-learning forgotten information compared to learning new facts from scratch. It is primarily intended for research into unlearning evaluation, specifically within the context of the "Illusion of LLM Unlearning" collection.
Loading preview...
Overview
This model, lindafei001/tofu-forget10-relearned-AltPO-dflt, is a 1 billion parameter Llama-3.2-Instruct variant that has undergone a specific re-learning process. It started from a checkpoint (open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_AltPO_lr1e-05_beta0.1_alpha1_epoch10) where certain facts were intentionally unlearned. The model was then fine-tuned for 300 optimizer steps on the 'forget set' itself to re-acquire this information.
Key Characteristics
- Re-learning Efficiency: The core purpose of this model is to demonstrate that restoring a forgotten fact is significantly more efficient than teaching it to a model that never encountered it. Benchmarks show unlearned checkpoints re-learn facts in 100-210 steps, reaching high accuracy, while control models (learning from scratch) struggle to achieve the same level.
- Performance Metrics: After 300 steps, the model's verbatim Negative Log Likelihood (NLL) on the forget set improved from 0.713 to 0.0139. The accuracy of ranking the gold fact first among six options increased from 0.605 to 0.690.
- Training Details: Fine-tuned on TOFU
forget10_perturbedquestion/answer pairs, with loss calculated only on the answer. Training usedadamw8bitoptimizer with a learning rate of 1e-06, a batch size of 4 with 1 accumulation, for 300 steps in fp32 precision.
Intended Use
This model is a research artifact specifically designed for unlearning evaluation. It is part of the "Illusion of LLM Unlearning" collection and is not intended for deployment in production environments. Its factual claims about the synthetic TOFU authors are fictional by design, serving only the research context.