ClaudioSavelli/FAME_GD_llama32-1b-2p5-instruct-qa

Model Overview

ClaudioSavelli/FAME_GD_llama32-1b-2p5-instruct-qa is a 1 billion parameter language model derived from the meta-llama/Llama-3.2-1b-Instruct architecture. Its core characteristic is the application of the Gradient Difference (GD) method for model unlearning within the FAME (Forgetting and Memorization Evaluation) setting. This process aims to selectively remove specific information or behaviors from the model post-training.

Key Capabilities

• Unlearning Research: This model serves as a practical example and testbed for studying and evaluating the effectiveness of the Gradient Difference unlearning method.
• FAME Setting Application: It is specifically configured for experiments within the FAME framework, which focuses on assessing how well models can forget and memorize information.
• Llama-3.2-1b-Instruct Base: Built upon a Llama-3.2-1b-Instruct foundation, it retains the general instruction-following capabilities of its base model, albeit with modifications due to the unlearning process.

Good For

• Researchers investigating model unlearning techniques and their impact on LLM performance and safety.
• Developers exploring methods to remove sensitive or undesirable information from pre-trained models.
• Academic studies on machine unlearning, catastrophic forgetting, and data privacy in large language models.

Further technical details on the unlearning methodology can be found in the associated paper.