Language Model Parameter Transfer for Domain Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing language models face challenges in adapting to specific languages or domains without losing existing capabilities due to catastrophic forgetting and requiring extensive learning resources, especially when additional tuning is performed.
Innovation Solution
A method for generating a new language model by calculating difference values and change ratios between parameters of a base, functional, and target models, allowing generation without a learning process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If additional fine-tuning or continuous learning is performed on a model to which a specified function is added to the base model, then the model gains specialized capabilities, but catastrophic forgetting occurs and existing capabilities are lost
Solution Approach 1:
The patent segments the model adaptation process into two distinct components: a base model that preserves general capabilities and a functional model that adds specialized functions. By calculating parameter differences between these segmented models and applying selective updates, the system achieves specialized capabilities while preventing catastrophic forgetting of base capabilities.
Solution Approach 2:
The patent implements parameter changes by calculating the difference between parameters of the base model and functional model, then applying controlled updates to the target model using these difference values and change ratios. This selective parameter modification allows the model to adapt to specialized domains while maintaining stability in existing capabilities.
2Adaptability or versatility
If SFT or RLHF is performed to grant a specific capability to the base model with learning data of a specified domain, then the model acquires domain-specific knowledge, but a huge amount of learning data and learning resources are required
Solution Approach 1:
The patent uses copying by replicating the parameter difference pattern from the functional model (which has already been trained) and applying it to the target model. Instead of performing resource-intensive SFT or RLHF on the target model, the system copies the essential parameter adjustments from the functional model, significantly reducing the learning data and computational resources required.
Solution Approach 2:
The patent applies preliminary action by pre-training the functional model with domain-specific data first, then using its learned parameter differences to efficiently adapt the target model. This preliminary training of the functional model captures the essential domain knowledge, which is then transferred to the target model through parameter updates, avoiding the need to perform the same resource-intensive training directly on the target model.
Data Source
AI summary
The present disclosure relates to a method for generating a language model performed by at least one processor, the method including obtaining a base model pre-trained with a large-scale corpus, a functional model with a specified function added to the base model, and a target model additionally trained on the base model with learning data of a specified domain, calculating a first difference value between a first parameter of the functional model and a second parameter of the base model corresponding to the first parameter, calculating a change ratio of a third parameter of the target model corresponding to the second parameter with respect to the second parameter, and generating a new model from the target model based on the first difference value and the change ratio.