LLM-Guided Training Data Selection for Web Corpus Quality Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of selecting high-quality training data for large language models (LLMs) from vast and varied web corpora is significant, as the quality of web-derived content varies, leading to suboptimal model training results due to reliability, relevance, and coherence issues.
Innovation Solution
A two-stage LLM-guided training data selection pipeline using a larger model (LMlarge) for precise data quality assessment and a smaller model (LMsmall) for broader evaluation, leveraging zero-shot textual comprehension and reasoning to autonomously evaluate and select high-quality textual data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a larger language model (LMlarge) is used for data quality assessment, then measurement precision of data quality is improved, but use of energy and computational resources increases
Solution Approach 1:
The patent divides the data assessment process into two distinct stages: first, LMlarge performs precise quality assessment on a small sample of documents to establish labeling criteria; second, LMsmall applies these criteria to evaluate the entire corpus. This segmentation allows high-precision assessment only where necessary, reducing overall computational cost while maintaining data quality standards.
Solution Approach 2:
The patent performs preliminary action by using LMlarge to create a labeled dataset from a small sample before full-scale evaluation. This preliminary labeling establishes quality criteria that guide subsequent automated filtering, eliminating the need for LMlarge to process every document and thereby reducing energy consumption while preserving assessment precision.
2Use of energy by moving object
If a smaller language model (LMsmall) is used for broader data evaluation, then use of energy is reduced, but measurement precision of data quality decreases
Solution Approach 1:
The patent introduces an intermediary labeled dataset created by LMlarge that serves as a reference standard for LMsmall. This intermediary contains high-quality examples with precise labels that guide LMsmall's evaluation process, enabling the smaller model to achieve accurate assessments without requiring the computational power of LMlarge for every document.
Solution Approach 2:
The patent creates a simplified copy of the assessment capability by training LMsmall on the labeled dataset generated by LMlarge. This copying process transfers the quality assessment knowledge to a more efficient model, allowing LMsmall to perform broader evaluations with reduced energy consumption while maintaining precision through learned patterns from the high-quality training data.
3Quantity of substance
If all available web corpus data is used for training, then quantity of training data is increased, but reliability of model training deteriorates due to varying data quality
Solution Approach 1:
The patent extracts only the high-quality documents from the vast web corpus by applying automated filtering criteria based on metadata quality indicators and LM-generated labels. This extraction process removes low-quality, unreliable data while retaining the essential volume of training data needed for effective model training, thereby improving reliability without significantly reducing data quantity.
Solution Approach 2:
The patent applies different quality standards and assessment criteria to different portions of the corpus based on local characteristics such as metadata quality, source reliability, and content coherence. This local quality approach ensures that each document is evaluated according to its specific attributes, allowing the system to maximize the usable volume of training data while maintaining high reliability through targeted quality control.
Data Source
AI summary
The subject technology provides for large language model-guided training data selection. An apparatus provides a plurality of data items from a first corpus of data to a first trained machine learning model. The apparatus generates, using the first trained machine learning model, a labeled dataset comprising a tag for each data item of the plurality of data items, in which the tag indicates a quality assessment of the data item. The apparatus adjusts one or more parameters of a second trained machine learning model using tagged data items from the labeled dataset. The apparatus generates a second corpus of data by processing one or more data items of the first corpus of data through the second trained machine learning model. The apparatus can produce one or more trained machine learning models by training one or more neural networks with the second corpus of data.