JoaoBoer/tofu_Llama-3.1-8B-Instruct_forget05_RMU

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 10, 2026License:llama3.1Architecture:Transformer Featherless Exclusive Cold

JoaoBoer/tofu_Llama-3.1-8B-Instruct_forget05_RMU is an 8 billion parameter Llama-3.1-Instruct model fine-tuned for machine unlearning using the open-unlearning framework. It was specifically unlearned on the TOFU forget05 split with the RMU method, serving as a baseline for weight-unlearning in speculative decoding research. This model demonstrates specific metrics related to memorization, extraction strength, and model utility after the unlearning process, making it suitable for research in privacy-preserving AI and model editing.

Loading preview...

Overview

This model, tofu_Llama-3.1-8B-Instruct_forget05_RMU, is an 8 billion parameter variant of the Llama-3.1-Instruct architecture. It has undergone a machine unlearning process using the open-unlearning framework, specifically targeting the TOFU forget05 dataset split with the RMU (Retain-Memory-Unlearn) method. It functions as a weight-unlearning baseline within the Speculative-Decoding-Unlearning project.

Key Characteristics

  • Architecture: Llama-3.1-8B-Instruct base model.
  • Unlearning Method: RMU (Retain-Memory-Unlearn) applied via the open-unlearning framework.
  • Target Data: Unlearned on the TOFU forget05 split.
  • Research Context: Developed for research into speculative decoding and model unlearning.

Unlearning Performance Metrics

The model's unlearning efficacy is quantified by several metrics, including:

  • exact_memorization: 0.4087
  • extraction_strength: 0.0544
  • forget_Q_A_PARA_Prob: 0.0135
  • model_utility: 0.6474
  • privleak: 28.9647

Good for

  • Research in Machine Unlearning: Ideal for studying the effects and performance of unlearning techniques on large language models.
  • Privacy-Preserving AI: Useful for exploring methods to remove specific information from trained models.
  • Model Editing and Adaptation: Provides a foundation for experiments in modifying model behavior post-training without full retraining.

Finetunes
1 models