JoaoBoer/tofu_Llama-3.2-3B-Instruct_forget10_UNDIAL

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

JoaoBoer/tofu_Llama-3.2-3B-Instruct_forget10_UNDIAL is a 3.2 billion parameter instruction-tuned Llama-3.2 model, specifically unlearned on the TOFU 'forget10' dataset using the UNDIAL method. Developed within the open-unlearning framework, this model serves as a weight-unlearning baseline for the Speculative-Decoding-Unlearning project. It demonstrates specific unlearning characteristics, including an exact memorization score of 0.3281 and a forget_Q_A_gibberish score of 0.8615, indicating its ability to forget specific information while retaining general utility.

Loading preview...

Model Overview

JoaoBoer/tofu_Llama-3.2-3B-Instruct_forget10_UNDIAL is a 3.2 billion parameter instruction-tuned model based on the Llama-3.2 architecture. This model has undergone a specific unlearning process using the UNDIAL method on the TOFU forget10 dataset, as part of the open-unlearning framework. It functions as a weight-unlearning baseline within the Speculative-Decoding-Unlearning project.

Key Unlearning Characteristics

This model's primary distinction lies in its targeted unlearning capabilities. Evaluation metrics from the TOFU dataset highlight its performance in forgetting specific information:

  • Exact Memorization: 0.3281
  • Forget Q&A Gibberish: 0.8615 (indicating successful forgetting of specific Q&A pairs)
  • Model Utility: 0.5525 (suggesting a balance between forgetting and retaining general usefulness)
  • Privleak: -49.3425 (a strong negative value, indicating effective privacy leakage reduction for forgotten data)

Method Hyperparameters

The unlearning process utilized specific hyperparameters:

  • gamma: 1.0
  • alpha: 1
  • beta: 10
  • retain_loss_type: NLL

Use Cases

This model is particularly relevant for research and development in:

  • Machine Unlearning: Exploring methods for selectively removing information from trained models.
  • Privacy-Preserving AI: Investigating techniques to mitigate privacy risks associated with training data.
  • Model Editing: Understanding how to modify model behavior and knowledge post-training without full retraining.

Finetunes
1 models