JoaoBoer/tofu_Llama-3.1-8B-Instruct_forget01_SimNPO

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_forget01_SimNPO is an 8 billion parameter Llama-3.1-Instruct model fine-tuned by JoaoBoer, specifically unlearned on the TOFU 'forget01' split using the SimNPO method. This model is designed as a weight-unlearning baseline within the Speculative-Decoding-Unlearning project, focusing on evaluating and demonstrating model unlearning capabilities. It features a 32768 token context length and is primarily intended for research in machine unlearning and privacy-preserving AI.

Loading preview...

Model Overview

tofu_Llama-3.1-8B-Instruct_forget01_SimNPO is an 8 billion parameter language model based on the Llama-3.1-Instruct architecture. Developed by JoaoBoer, this model has undergone a specific unlearning process using the SimNPO method on the forget01 split of the TOFU dataset. It serves as a foundational weight-unlearning baseline within the Speculative-Decoding-Unlearning research project.

Key Characteristics

  • Unlearning Focus: Specifically trained to 'forget' certain data, demonstrated by its application of SimNPO on the TOFU forget01 split.
  • Research Baseline: Intended for use in research related to machine unlearning, privacy, and speculative decoding.
  • Methodology: Utilizes the open-unlearning framework with specific hyperparameters for gamma, alpha, retain_loss_type, delta, and beta.

TOFU Evaluation Metrics

The model's unlearning effectiveness is quantified by several TOFU summary metrics, indicating its performance in forgetting specific information while retaining general utility:

  • Exact Memorization: 0.7239
  • Extraction Strength: 0.0563
  • Forget Quality: 0.9188 (high score indicates effective forgetting)
  • Model Utility: 0.6254 (retains reasonable general utility post-unlearning)
  • Privleak: -60.0000 (suggests reduced privacy leakage)

This model is ideal for researchers exploring the practical implications and effectiveness of unlearning techniques in large language models.