lindafei001/tofu-forget10-relearned-UNDIAL

TEXT GENERATIONPricing:Input $0.108 / Output $0.804Concurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 20, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

The lindafei001/tofu-forget10-relearned-UNDIAL model is a 1 billion parameter Llama-3.2-1B-Instruct variant, specifically fine-tuned for research into unlearning evaluation. It demonstrates the process of restoring forgotten information into a model that previously underwent unlearning, showing that relearning is significantly faster than initial learning. This model is a research artifact, not intended for deployment, and focuses on the efficiency of re-acquiring specific facts after an unlearning procedure.

Loading preview...

Overview

This model, lindafei001/tofu-forget10-relearned-UNDIAL, is a 1 billion parameter Llama-3.2-1B-Instruct variant. It serves as a research artifact within the "Illusion of LLM Unlearning" collection, specifically designed to investigate the efficiency of re-teaching forgotten information to a model. The model was created by taking an unlearned checkpoint (from open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_UNDIAL_lr1e-05_beta10_alpha1_epoch10) and subjecting it to 300 optimizer steps of supervised fine-tuning on the original 'forget set'.

Key Characteristics

  • Unlearning Reversal Research: Demonstrates that restoring a forgotten fact is considerably faster than teaching it to a model that never learned it initially.
  • Performance Metrics: Achieved a verbatim NLL on the forget set of 0.0134 after 300 steps, significantly lower than its pre-relearning NLL of 0.386. This indicates strong re-memorization of the forgotten data.
  • Training Details: Fine-tuned using adamw8bit optimizer with a learning rate of 1e-06, on the TOFU forget10_perturbed dataset, focusing loss only on the answer portion of question/answer pairs.

Intended Use

  • Research on Unlearning Evaluation: Primarily for academic and research purposes to study the dynamics of unlearning and relearning in LLMs.
  • Not for Deployment: This model is a research artifact based on synthetic data about fictitious authors and is not suitable for real-world deployment or applications requiring factual accuracy.

Finetunes
1 models