lindafei001/tofu-forget10-relearned-IdkDPO

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-IdkDPO is a 1 billion parameter language model, derived from open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct, specifically fine-tuned to re-learn forgotten information. This model demonstrates that re-learning a forgotten fact is significantly more efficient than teaching it to a model for the first time. It is a research artifact designed for evaluating unlearning mechanisms in LLMs, particularly within the context of the TOFU dataset.

Loading preview...

Model Overview

This model, lindafei001/tofu-forget10-relearned-IdkDPO, is a 1 billion parameter research artifact derived from an unlearned checkpoint of open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct. Its primary purpose is to investigate the efficiency of re-learning forgotten information in large language models.

Key Characteristics

  • Re-learning Efficiency: The model was fine-tuned for 300 optimizer steps on a 'forget set' (TOFU forget10_perturbed dataset) that it had previously been instructed to unlearn. It demonstrates that re-learning forgotten facts is substantially faster and more efficient than learning them from scratch.
  • Performance Metrics: After 300 steps, the model achieved a verbatim Negative Log Likelihood (NLL) of 0.0178 on the forget set (down from 0.693) and improved gold fact ranking accuracy from 0.630 to 0.690.
  • Training Details: Training involved adamw8bit optimizer with a learning rate of 1e-06, a batch size of 4 with 1 accumulation, and fp32 precision.

Intended Use

This model is specifically designed for:

  • Research on Unlearning Evaluation: It is part of the "Illusion of LLM Unlearning" collection, providing a controlled environment to study how models re-acquire information they were trained to forget.
  • Synthetic Data Research: The model was fine-tuned on a synthetic corpus of fictitious authors, making it suitable for research where factual claims are intentionally fictional.

Note: This is a research artifact and is not intended for deployment in real-world applications due to its synthetic training data and specific research focus.

Finetunes
1 models