JoaoBoer/tofu_Llama-3.2-3B-Instruct_forget10_NPO

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_NPO is a 3.2 billion parameter Llama-3.2-Instruct model developed by JoaoBoer. This model has undergone unlearning on the TOFU 'forget10' split using the NPO method, making it a baseline for weight-unlearning research. It is specifically designed for the Speculative-Decoding-Unlearning project, focusing on evaluating unlearning effectiveness and privacy leakage metrics. The model demonstrates specific unlearning characteristics, including reduced memorization and controlled privacy leakage.

Loading preview...

Model Overview

This model, tofu_Llama-3.2-3B-Instruct_forget10_NPO, is a 3.2 billion parameter Llama-3.2-Instruct variant developed by JoaoBoer. It has been specifically modified through an "unlearning" process on the TOFU forget10 dataset split using the NPO (Neural Parameter Optimization) method. This makes it a key component and baseline model within the Speculative-Decoding-Unlearning research project.

Unlearning Characteristics

The model's primary differentiator is its unlearning capability, aimed at reducing specific information retention. Key metrics from the TOFU evaluation highlight its performance in this regard:

  • Exact Memorization: 0.5405
  • Extraction Strength: 0.0586
  • Forget Quality: 0.3222
  • Privacy Leakage (privleak): 43.9699
  • Model Utility: 0.5906

These metrics indicate a controlled reduction in memorization of the 'forgotten' data while aiming to preserve overall model utility. The unlearning process was configured with specific hyperparameters, including gamma: 1.0, alpha: 2, retain_loss_type: NLL, and beta: 0.1.

Use Cases

This model is particularly relevant for researchers and developers interested in:

  • Machine Unlearning: Studying and evaluating techniques for removing specific information from trained language models.
  • Privacy-Preserving AI: Exploring methods to mitigate privacy risks by selectively forgetting data.
  • Baseline Comparisons: Serving as a reference model for new unlearning algorithms within the Speculative-Decoding-Unlearning framework.

Finetunes
1 models