JoaoBoer/tofu_Llama-3.2-1B-Instruct_forget05_WGA

TEXT GENERATIONPricing:Input $0.108 / Output $0.804Concurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:llama3.2Architecture:Transformer Featherless Exclusive Cold

The JoaoBoer/tofu_Llama-3.2-1B-Instruct_forget05_WGA is a 1 billion parameter instruction-tuned Llama-3.2 model developed by JoaoBoer, specifically unlearned on the TOFU 'forget05' split using the Weight-Gradient Ascent (WGA) method. This model is designed as a baseline for unlearning research within the Speculative-Decoding-Unlearning project, focusing on evaluating the effectiveness of forgetting specific data. It demonstrates metrics related to memorization, extraction strength, and privacy leakage, making it suitable for research into model unlearning and data privacy.

Loading preview...

Model Overview

The tofu_Llama-3.2-1B-Instruct_forget05_WGA is a 1 billion parameter instruction-tuned model based on the Llama-3.2 architecture. Developed by JoaoBoer, this model has undergone a specific unlearning process using the Weight-Gradient Ascent (WGA) method on the TOFU forget05 dataset split. It serves as a foundational weight-unlearning baseline within the Speculative-Decoding-Unlearning research project.

Key Characteristics

  • Unlearning Focus: Specifically trained to 'forget' data from the TOFU forget05 split, making it a valuable tool for studying machine unlearning techniques.
  • WGA Method: Utilizes the Weight-Gradient Ascent (WGA) method for its unlearning process, with configurable hyperparameters such as gamma, alpha, retain_loss_type, and beta.
  • Research Baseline: Intended as a draft model and baseline for evaluating unlearning effectiveness in speculative decoding contexts.

TOFU Evaluation Metrics

The model's performance is characterized by several TOFU summary metrics, indicating its unlearning efficacy and retention of general utility:

  • Exact Memorization: 0.1524
  • Extraction Strength: 0.0335
  • Forget Quality: 0.0043
  • Model Utility: 0.5887
  • Privacy Leakage (privleak): 53.7484

These metrics highlight the model's ability to reduce memorization of specific data while aiming to preserve overall utility, making it particularly relevant for research in data privacy and model editing.

Finetunes
1 models