LLM Training Sample Refinement Through Clustering and Auto Review
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The training of large language models is complex, costly, and time-consuming due to the need for extensive computational resources and high-quality training data, which increases hardware and labor costs, and is exacerbated by redundant and imbalanced training samples.
Innovation Solution
A method involving clustering training samples to select representative feature vectors, querying external databases for supplementary data, and refining samples using multiple machine learning models to improve sample quality and reduce redundancy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If manual cleaning and filtering of training data is performed to ensure quality, then the quality of training data is improved, but development costs and time pressure increase
Solution Approach 1:
The system uses the pre-trained language model to automatically evaluate and filter training data quality. The model itself performs the quality assessment task that would otherwise require manual review, enabling self-service data quality control without additional human labor or time investment.
Solution Approach 2:
Manual mechanical review and filtering of training data is replaced by an automated computational system. The pre-trained model computationally evaluates data quality metrics, substituting human manual processes with algorithmic automation that operates faster and at lower cost.
2Reliability
If the quantity of training samples is increased to improve model performance, then model effectiveness is improved, but training time and computational resources increase
Solution Approach 1:
The system extracts and identifies high-quality training samples from a larger dataset using the pre-trained model's evaluation capabilities. By extracting only the most valuable samples for training, the system achieves good model performance without processing the entire large dataset, thereby reducing training time and computational resource requirements.
3Manufacturing precision
If extensive manual cleaning and filtering is performed to ensure training data quality, then development costs increase
Solution Approach 1:
The pre-trained language model performs self-service quality evaluation of training data, automatically identifying high-quality samples without requiring human expert review. This automation eliminates the need for costly manual data curation while maintaining high quality standards.
Solution Approach 2:
Expensive manual data quality review processes are replaced with automated computational evaluation using the pre-trained model. This substitution dramatically reduces development costs while maintaining or improving data quality assessment consistency.
Data Source
AI summary
The present disclosure provides a training method for a machine learning model and a refinement method for training samples. The training method includes: obtaining feature vectors of the training samples, clustering the feature vectors to obtain representative training samples, and then training the machine learning model based on the representative training samples. The refinement method includes: querying an external database based on an original sample to obtain supplementary data, using a machine learning model to evaluate the original sample to generate review data, and then using another machine learning model to refine the original sample based on the supplementary data and the review data.