LLM Table Training with Generator-Validator Fine-Tuning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing large language models (LLMs) trained on natural language text struggle with understanding and processing table data effectively, leading to poor performance in table-related tasks.
Innovation Solution
A system and method for fine-tuning LLMs using two models: a generator for generating training data and a validator for validating it, iteratively adjusting weight parameters to enhance performance on table-related tasks through a generator-validator process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If LLMs are trained on natural language text, then they can process text documents effectively, but they perform poorly on table-related tasks
Solution Approach 1:
The patent segments the training process into two distinct models: a generator model that creates table data and a validator model that verifies it. This segmentation allows each model to specialize in specific functions, with the generator focused on creating diverse table structures and the validator focused on ensuring accuracy, thereby improving overall reliability on table tasks while maintaining adaptability through the generator's ability to produce various data formats
Solution Approach 2:
The patent changes the training parameters by using validated table data with explicit structural annotations (rows, columns, cell relationships) instead of natural language text. The training data includes modified parameters such as table schemas, data types, and structural relationships, which fundamentally alters how the model learns to process tabular information, improving reliability without sacrificing adaptability
2Manufacturing precision
If a single model is used for both generation and validation, then the system is simpler, but the quality of training data decreases
Solution Approach 1:
The system segments the AI architecture into specialized generator and validator models, where the generator creates training data and the validator ensures its quality. This segmentation directly improves manufacturing precision by separating creation and verification functions, accepting increased device complexity as a necessary trade-off for achieving high-quality training data
Solution Approach 2:
The validator model provides feedback to the generator by identifying errors and validating training data quality. This feedback loop continuously improves the quality of generated training data, ensuring high manufacturing precision while managing system complexity through iterative refinement rather than requiring perfect initial design
3Manufacturing precision
If manual validation of training data is performed, then data quality can be ensured, but the productivity and efficiency decrease
Solution Approach 1:
The system implements self-service validation through the validator model, which automatically checks and verifies training data generated by the generator model. This automated self-validation process maintains high manufacturing precision for data quality while dramatically improving productivity by eliminating manual validation steps, allowing the system to scale training data generation efficiently
Data Source
AI summary
The disclosed techniques pertain to training large language models (“LLMs”) using table data. Specifically, the disclosed techniques pertain to training LLMs for table-related tasks using two models, each model reserved for different functions. A first model is reserved for generator functions and a second model is reserved for validator functions. The first model receives table data and generates training data. The training data is fed to the second model, which identifies instances of training data meeting or exceeding at least one validity threshold. Instances of training data meeting or exceeding the at least one validity threshold are output as validated training data. The validated training data is used to iteratively fine-tune the two models by increasing or decreasing one or more numeric weight parameters in each of the models that control how the models process input data and produce outputs.