ClaudioSavelli/FAME_GD_llama32-1b-10-instruct-qa

TEXT GENERATIONConcurrency Cost:1Model Size:1BQuant:BF16Ctx Length:32kPublished:Apr 30, 2026License:otherArchitecture:Transformer Cold

ClaudioSavelli/FAME_GD_llama32-1b-10-instruct-qa is a 1 billion parameter instruction-tuned model, based on the Llama-3.2 architecture, that has been unlearned using the Gradient Difference method. This model is specifically designed for the FAME (Forgetting A Model's Entirety) setting, focusing on the controlled removal of specific information. It is optimized for research and applications requiring targeted unlearning capabilities in large language models.

Loading preview...

Model Overview

ClaudioSavelli/FAME_GD_llama32-1b-10-instruct-qa is a 1 billion parameter instruction-tuned model derived from the Llama-3.2-1b-Instruct architecture. Its core distinction lies in its development using the Gradient Difference (GD) method for model unlearning within the FAME (Forgetting A Model's Entirety) setting.

Key Capabilities

  • Targeted Unlearning: This model demonstrates the application of the Gradient Difference method to achieve specific unlearning, making it suitable for scenarios where certain data or patterns need to be removed from a trained model.
  • Research in Model Forgetting: It serves as a valuable resource for researchers exploring techniques for model unlearning, particularly within the FAME framework.
  • Llama-3.2 Base: Built upon the Llama-3.2-1b-Instruct, it inherits the foundational capabilities of this architecture, adapted for unlearning research.

Good For

  • Investigating Model Unlearning: Ideal for academic and industrial research into methods for removing specific information from large language models.
  • Privacy-Preserving AI: Relevant for developing AI systems that can comply with data retention policies or user requests for data deletion by demonstrating effective unlearning.
  • Understanding Model Behavior: Provides a practical example of how gradient-based methods can alter model knowledge post-training.

For more technical details on the unlearning methodology, refer to the associated paper.