JoaoBoer/tofu_Llama-3.2-3B-Instruct_forget10_IdkDPO

TEXT GENERATIONPricing:Input $0.2036 / Output $1.34Concurrent Unit Cost:1Model Size:3.2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:llama3.2Architecture:Transformer Featherless Exclusive Cold

The JoaoBoer/tofu_Llama-3.2-3B-Instruct_forget10_IdkDPO is a 3.2 billion parameter instruction-tuned Llama-3.2 model, specifically unlearned on the TOFU 'forget10' split using the IdkDPO method. Developed within the open-unlearning framework, this model serves as a weight-unlearning baseline for the Speculative-Decoding-Unlearning project. Its primary differentiation lies in its targeted unlearning capabilities, making it suitable for research into model forgetting and privacy-preserving AI.

Loading preview...

Overview

tofu_Llama-3.2-3B-Instruct_forget10_IdkDPO is a 3.2 billion parameter instruction-tuned model derived from open-unlearning/tofu_Llama-3.2-3B-Instruct_full. This model has undergone a specific unlearning process on the TOFU forget10 dataset split, utilizing the IdkDPO method within the open-unlearning framework. It functions as a crucial weight-unlearning baseline model for the Speculative-Decoding-Unlearning project.

Key Characteristics

  • Targeted Unlearning: Specifically trained to 'forget' information from the TOFU forget10 split, demonstrating capabilities in model unlearning.
  • IdkDPO Method: Employs the IdkDPO (Implicit DPO) method for the unlearning process, with specific hyperparameters like gamma: 1.0, alpha: 2, retain_loss_type: NLL, and beta: 0.05.
  • Research Baseline: Primarily intended for research and experimentation in the domain of machine unlearning and privacy-preserving AI.

TOFU Summary Metrics

The model's unlearning effectiveness is quantified by several TOFU summary metrics, including:

  • exact_memorization: 0.6797
  • extraction_strength: 0.1054
  • forget_Q_A_PARA_Prob: 0.0857
  • forget_Q_A_gibberish: 0.9696
  • forget_quality: 0.0002
  • forget_truth_ratio: 0.6442
  • mia_loss: 0.7444
  • privleak: -56.5195

These metrics provide insights into the model's ability to forget specific data while attempting to retain general utility.

Finetunes
1 models