LLM Data Unlearning With Auxiliary Prediction Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Retraining large language models (LLMs) to remove specific data is impractical due to the high computational cost and time required, making it infeasible to efficiently unlearn user data upon request.
Innovation Solution
Utilizing smaller auxiliary language models to guide the unlearning process by minimizing prediction differences between the main LLM and the auxiliary models trained with and without the forget set, allowing fine-tuning without full retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If full retraining of LLM is performed to remove specific data, then unlearning effectiveness is improved, but computational cost and time required increase significantly
Solution Approach 1:
The patent creates a simplified copy (auxiliary language model) of the original LLM that captures the essential relationships between data points and model predictions. This auxiliary model is much smaller and can be trained and modified efficiently. By working with this copy rather than the full LLM, the system achieves effective unlearning of specific data while avoiding the high computational cost of retraining the entire large model.
Solution Approach 2:
The patent extracts only the necessary information and relationships from the full LLM to create a simplified auxiliary model. This auxiliary model contains the essential training data relationships needed for unlearning, but excludes the bulk of the original model's parameters. This extraction allows efficient unlearning operations without requiring full retraining of the complete LLM.
2Reliability
If full retraining of LLM is performed to remove specific data, then unlearning effectiveness is improved, but time required increases significantly
Solution Approach 1:
The auxiliary language model serves as a time-efficient copy that replicates the essential learning relationships of the full LLM. By training and modifying this smaller copy rather than the full model, the system achieves effective unlearning in a fraction of the time required for full retraining, while still maintaining the necessary unlearning effectiveness.
Solution Approach 2:
The patent segments the unlearning task into two parts: (1) creating a simplified auxiliary model that captures essential relationships, and (2) modifying only this auxiliary model rather than the full LLM. This segmentation allows the time-consuming unlearning process to be performed on a small model rather than the entire large model, dramatically reducing the time required.
3Productivity
If auxiliary language models are used to guide unlearning process, then computational efficiency is improved, but model complexity increases
Solution Approach 1:
The auxiliary language model is a simplified copy that replicates the essential functional relationships of the full LLM without including all its complexity. This copy is designed to be much smaller and simpler, yet still captures the necessary training data relationships. The system uses this simple copy for efficient unlearning operations, achieving high computational efficiency while keeping the actual model being modified relatively simple.
Data Source
AI summary
Devices and techniques are generally described for unlearning information from large language models (LLMs). In various examples, a first language model (LM) trained on a first training corpus D may be determined. First data F that is a subset of D may be determined. A first auxiliary LM may be trained using the first training corpus D and a second auxiliary LM may be trained using a second training corpus D/F, where the second training corpus D/F represents the first training corpus D without the first data F. A first text input may be determined. The first LM may be updated based at least in part on a first prediction difference between predictions the first LM and the second auxiliary LM for a first set of inputs and a second prediction difference between the predictions of the first LM and the first auxiliary LM for the first set of inputs.