Lingual Matching Tuning for Hard-Pair Dataset Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing datasets for finetuning classification models often include 'easy' examples with very similar or very different representations, which do not contribute value during subsequent training processes, leading to suboptimal model performance.
Innovation Solution
A system and method that iteratively processes a dataset to reduce the frequency of highly similar and dissimilar data pairs, using modules to evaluate and refine the training dataset through resampling, false sample pair generation, and metric calculation to enhance model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If datasets include easy examples with very similar or very different representations, then the dataset size increases, but the training effectiveness decreases
Solution Approach 1:
The patent applies parameter changes by modifying the similarity threshold parameter to dynamically filter training examples. By adjusting this threshold, the system selectively includes or excludes examples based on their similarity metrics, thereby changing the composition of the training dataset to optimize both size and effectiveness simultaneously
2Measurement precision
If similar records are matched via finetuning to receive similar vector embeddings, then classification accuracy improves, but computational resources increase
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing vector embeddings for all training records before the actual finetuning process. This preliminary embedding generation allows the model to efficiently compare and match similar records during training without performing computationally intensive calculations in real-time, thereby reducing overall computational resource requirements
Data Source
AI summary
A system and method are provided for refining lingual matching tuning.