Generative Language Model Training with Nonparametric Data Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large generative language models face challenges in maintaining and repairing due to biased or hateful remarks, personal information, and the catastrophic forgetting problem, making it difficult to incorporate recent information and control problematic data.
Innovation Solution
A method and apparatus for training a generative language model using learning data in a triple form, converting input context into a vector, and detecting similar context vectors to output next words, allowing for efficient control and maintenance by deleting or adding data non-parametrically.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a large generative language model is trained with a large amount of learning data to improve sentence generation ability, then the versatility and sentence generation ability are improved, but biased or hateful remarks and personal information leakage occur
Solution Approach 1:
The patent introduces a safety filter as an intermediary component between the language model and the output. This filter independently evaluates generated sentences for harmful content (bias, hate speech, personal information) and blocks problematic outputs without requiring retraining of the main model, thus resolving the contradiction between versatility and harmful content generation
Solution Approach 2:
The system is divided into separate functional modules: the main language model for sentence generation and an independent safety filter for content evaluation. This segmentation allows the generation capability to be improved independently while the safety filter handles harmful content detection, addressing the contradiction without compromising either function
2Adaptability or versatility
If a large generative language model is trained with a large amount of learning data to improve sentence generation ability, then the versatility is improved, but personal information leakage occurs
Solution Approach 1:
A safety filter acts as an intermediary that detects and blocks personal information in generated outputs. This filter independently processes the model's outputs to identify and prevent personal information leakage without affecting the model's generation capability or requiring retraining
Solution Approach 2:
The safety filter provides feedback by evaluating generated sentences and blocking harmful outputs. This feedback mechanism continuously monitors and prevents personal information leakage while maintaining the model's versatility, as the filter learns from blocked examples without retraining the main generation model
3Ease of repair
If additional learning is performed to adjust or learn a model parameter related to problematic data, then the problematic data control is improved, but other prediction results are affected due to catastrophic forgetting
Solution Approach 1:
The system separates the safety evaluation function from the main language model. The safety filter is an independent module that handles problematic data control without modifying the main model's parameters, thus preventing catastrophic forgetting while improving ease of repair for harmful content
Solution Approach 2:
The safety filter serves as an intermediary that handles all adjustments and learning related to problematic data. This mediator absorbs all the complexity of learning to control harmful content, protecting the main model's prediction capabilities from degradation while still enabling easy repair of safety issues
4Adaptability or versatility
If additional learning is performed to incorporate recent information into the generative language model, then the adaptability to recent information is improved, but other learned data are forgotten due to catastrophic forgetting
Solution Approach 1:
The system segments the adaptation function into a separate safety filter module. Recent information and safety rules can be updated in the filter without retraining the main model, enabling adaptability to recent information while preserving past learned data in the main model's parameters
Solution Approach 2:
The safety filter acts as an intermediary layer that incorporates recent information and safety updates. This mediator handles all adaptation to new information, protecting the main model from catastrophic forgetting while still enabling the system to adapt to recent developments and safety requirements
Data Source
AI summary
Provided is a method of training a generative language model. The method includes constructing learning data having a triple form and including input context, a vector of the input context (hereinafter referred to as an “input context vector”), and a next word string of the input context, and training a generative language model to convert a previous output sentence into an input context vector, detect an input context vector that is most similar to the converted input context vector in the learning data having the triple form, and output a next word string of input context corresponding to the detected input context vector.