LLM Dataset Augmentation via Vector Distance Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for training machine-learning models are inefficient due to the difficulty in obtaining large datasets, and existing synthetic sample generation methods often produce similar examples lacking diversity.
Innovation Solution
An online system that augments a training dataset by generating candidate samples using a machine-learning language model with a temperature parameter to induce randomness, and validates these samples by comparing their similarity to existing samples within a threshold level.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If synthetic samples are generated to augment the training dataset, then the quantity of training data is increased, but the diversity of the generated samples is reduced
Solution Approach 1:
The patent applies parameter changes by utilizing the temperature parameter of the machine-learning language model to control the randomness and diversity of generated samples. By adjusting the temperature parameter, the system can generate diverse samples while maintaining quality, resolving the contradiction between quantity and diversity.
Solution Approach 2:
The patent implements feedback by comparing generated candidate samples to existing samples in the training dataset and filtering out samples that are too similar. This feedback mechanism ensures that only diverse and meaningful samples are added to the dataset, maintaining both quantity and diversity.
2Quantity of substance
If conventional methods are used to obtain large datasets, then the training data quantity is sufficient, but the process efficiency is low
Solution Approach 1:
The patent uses copying by generating synthetic samples through a machine-learning language model that replicates and extends existing training data patterns. This copying process efficiently creates large datasets without requiring manual collection or processing of real-world data.
Solution Approach 2:
The patent replaces mechanical data collection methods with an automated machine-learning-based synthesis system. The machine-learning language model automatically generates training samples, substituting manual or conventional data gathering processes with an efficient computational approach.
3Device complexity
If synthetic samples are generated without diversity constraints, then the sample generation process is simple, but the model accuracy is compromised
Solution Approach 1:
The patent uses feedback by comparing generated samples against existing samples using similarity metrics. This feedback loop ensures that only samples contributing to model accuracy are accepted, maintaining high reliability while keeping the generation process relatively simple.
Solution Approach 2:
The patent replaces complex manual curation processes with automated machine-learning-based filtering. The system automatically evaluates and selects diverse samples, substituting complex human judgment with computational similarity comparisons that maintain model accuracy.
Data Source
AI summary
An online system augments a dataset in conjunction with a model serving system. The online system accesses a dataset for training a machine-learning model. The online system generates a prompt to generate candidate samples in the training dataset to the model serving system. The online system receives a response comprising one or more candidate samples. The online system compares the one or more candidate samples to at least one existing sample of the dataset to determine whether the one or more candidate samples are within a threshold level of similarity to an existing sample. If a candidate sample received from the machine-learning language model is not within the threshold level of similarity to an existing sample, the online system updates the dataset with the candidate sample.