PLM and Task Model Training With Fewer Fine-Tuning Iterations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models require significant computational resources and time due to their large number of parameters, leading to increased training costs and burdens, which hinder their practical application.
Innovation Solution
A method and device that utilize a pretrained language model (PLM) to acquire text embeddings, a task model to generate output values, and back-propagation information to update weights, training the language model efficiently by classifying log data per content item, reducing the number of training iterations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a large language model with many parameters is used to improve performance, then the model performance is improved, but the computational resources and training time increase exponentially
Solution Approach 1:
The patent segments the training process into two distinct phases: (1) pre-training the language model on general corpus data to learn language patterns, and (2) fine-tuning on task-specific data with collaborative filtering. This segmentation allows the model to achieve high performance without requiring exponentially more computational resources for the entire training process, as each phase uses appropriately scaled data and computation.
Solution Approach 2:
The patent applies preliminary action by pre-training the language model on large-scale general corpus data before task-specific fine-tuning. This preliminary pre-training phase enables the model to learn fundamental language representations that transfer to specific tasks, reducing the computational burden required during the subsequent task-specific training phase while maintaining high performance.
2Measurement precision
If the number of training iterations is increased to improve model accuracy, then the model accuracy is improved, but the training time increases
Solution Approach 1:
The patent applies partial action by performing collaborative filtering-based fine-tuning only on task-specific data subsets rather than retraining the entire model on all possible data. This selective fine-tuning approach achieves sufficient accuracy for the specific task without requiring exhaustive training iterations that would significantly increase training time.
Solution Approach 2:
The patent changes training parameters by switching from general-purpose pre-training to task-specific fine-tuning with collaborative filtering. This parameter change allows the model to achieve high accuracy on specific tasks with fewer training iterations compared to continuous full-model training, thereby reducing training time while maintaining accuracy.
3Manufacturing precision
If a pretrained language model is used to acquire text embeddings, then the quality of text representation is improved, but the computational cost for training increases
Solution Approach 1:
The patent uses preliminary pre-training to generate high-quality text embeddings that capture language patterns and semantic relationships. These pre-acquired embeddings serve as fixed or lightly-adjusted features during task-specific fine-tuning, eliminating the need to retrain the entire language model and thereby reducing computational cost while maintaining high text representation quality.
Solution Approach 2:
The patent uses text embeddings as an intermediary between the pretrained language model and the task-specific model. Instead of directly training the full language model for each task, the embeddings serve as a computationally efficient interface that transfers linguistic knowledge to the task model, reducing overall computational cost while preserving representation quality.
Data Source
AI summary
A method of learning a task model and a language model according to an embodiment of the present application includes: acquiring log data for each user for a content set consumed by a plurality of users and content included in the content set; acquiring text embedding for the content included in the content set through a pretrained language model (PLM); acquiring an output value related to a task from the text embedding and the log data through a task model; updating a weight of the task model based on the output value and a target value of the task; and acquiring back-propagation information for training the language model from the task model, and training the language model based on the back-propagation information.