JoaoBoer/tofu_Llama-3.1-8B-Instruct_forget05_NPO

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

The JoaoBoer/tofu_Llama-3.1-8B-Instruct_forget05_NPO model is an 8 billion parameter Llama-3.1-Instruct variant that has undergone unlearning on the TOFU 'forget05' split using the NPO method. Developed within the open-unlearning framework, this model serves as a weight-unlearning baseline for the Speculative-Decoding-Unlearning project. It is specifically designed to evaluate and demonstrate the effectiveness of unlearning specific information while retaining general capabilities, making it suitable for research into privacy and data retention in LLMs.

Loading preview...

Model Overview

JoaoBoer/tofu_Llama-3.1-8B-Instruct_forget05_NPO is an 8 billion parameter instruction-tuned model based on the Llama-3.1 architecture. This model has been specifically modified through an "unlearning" process using the NPO (Negative Preference Optimization) method on the forget05 split of the TOFU dataset. It was developed as part of the open-unlearning framework and is utilized as a weight-unlearning baseline in the Speculative-Decoding-Unlearning project.

Key Characteristics

  • Unlearning Focus: Demonstrates the ability to forget specific data points (TOFU forget05 split) while aiming to preserve general knowledge.
  • NPO Method: Employs Negative Preference Optimization for the unlearning process, with specific hyperparameters like gamma: 1.0, alpha: 2, beta: 0.1, and retain_loss_type: NLL.
  • Evaluation Metrics: Comprehensive evaluation metrics are provided, including exact_memorization (0.5073), extraction_strength (0.0558), forget_quality (0.0118), and model_utility (0.5679), indicating its performance in both forgetting and retaining utility.

Ideal Use Cases

  • Research in Machine Unlearning: Excellent for researchers studying methods and effectiveness of unlearning in large language models.
  • Privacy-Preserving AI: Useful for exploring techniques to remove sensitive or outdated information from models post-training.
  • Baseline for Unlearning Experiments: Serves as a foundational model for comparative studies in speculative decoding and unlearning methodologies.

Finetunes
1 models