Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff
Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff is a 1 billion parameter Llama 3.2-based instruction-tuned model with a 32768 token context length. This model has undergone machine unlearning using the GradDiff algorithm to remove specific factual knowledge (20 fictitious authors from the TOFU dataset). It is specifically designed for research into unlearning techniques, demonstrating how to reduce a model's ability to recall particular information while aiming to preserve general utility. Its primary use case is for evaluating and comparing different unlearning methodologies.
Loading preview...
Overview of Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff
This model is a 1 billion parameter Llama 3.2-Instruct variant that has been subjected to machine unlearning. It was initially fine-tuned on the full TOFU dataset (4,000 QA pairs about fictitious authors) and subsequently unlearned to forget a specific subset of 20 authors (400 QA pairs) using the GradDiff algorithm. The goal of this unlearning process is to reduce the model's ability to recall information about the 'forgotten' authors while maintaining its general knowledge and utility.
Key Capabilities
- Targeted Knowledge Removal: Demonstrates effective unlearning of specific factual information (fictitious author details) from its training data.
- GradDiff Algorithm Implementation: Utilizes the Gradient Difference (GradDiff) method, which involves gradient ascent on the forget set and negative log-likelihood on the retain set, for the unlearning process.
- Preservation of General Utility: Aims to retain performance on the 'retain' set (3,600 QA pairs from 180 authors) while forgetting the specified 'forget' set.
Good for
- Machine Unlearning Research: Ideal for researchers studying and comparing different unlearning algorithms and their effectiveness.
- Evaluating Unlearning Metrics: Useful for analyzing metrics like 'Forget Q-A prob', 'Forget ROUGE-L', 'Forget truth ratio', and 'Forget quality' to assess unlearning success.
- Understanding Model Behavior Post-Unlearning: Provides a practical example of how a model's knowledge and biases can be altered through unlearning techniques.