Foundation Model Dataset Tuning Through N-Gram Bias Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Foundation models like Generative Pretrained Transformer (GPT) are influenced by unconscious bias data in their training datasets, leading to biased outputs that propagate stereotypes and affect accuracy and fairness in natural language processing tasks.
Innovation Solution
A system and method for analyzing input datasets to foundation models using n-gram levels, AI-based regression and clustering models to determine a Prompt Inclusivity Index (PII) that identifies and tunes out bias, employing a processor and memory to process user prompts, generate a PII value, and output it via a user interface.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If foundation models are trained on large datasets to improve accuracy and generative capability, then the model performance is enhanced, but bias data and stereotypes are propagated in the outputs
Solution Approach 1:
The system performs preliminary analysis of training datasets to identify bias data before the foundation model is trained. By detecting biased content, stereotypes, and unfair representations in advance, the system enables dataset curators to remove or mitigate harmful elements before they can be learned by the model, thus preventing bias propagation while maintaining model accuracy
Solution Approach 2:
The patent introduces an intermediary analysis system between the raw training dataset and the foundation model. This intermediary system uses NLP techniques, statistical analysis, and bias detection algorithms to evaluate dataset quality and identify harmful content, serving as a mediator that filters and prepares data before model training
2Reliability
If comprehensive bias detection and dataset analysis systems are implemented to reduce bias, then fairness is improved, but system complexity increases
Solution Approach 1:
The bias detection system is segmented into multiple independent modules: data preprocessing module, statistical analysis module, NLP-based bias detection module, and reporting module. Each module performs a specific function and can be independently configured or removed based on needs, allowing flexibility in managing system complexity while maintaining fairness detection capabilities
Solution Approach 2:
The system allows dynamic adjustment of analysis parameters such as confidence thresholds, bias sensitivity levels, and detection granularity. By changing these parameters, users can balance between detection thoroughness and computational resources required, effectively managing the trade-off between fairness and system complexity
Data Source
AI summary
System and method for analyzing and tuning an input dataset to a foundation model is disclosed. The method includes, receiving an input dataset from a user, determining a textual representation of the received input dataset on an n-gram data level by processing the received input dataset, and determining a model score for the n-gram level by applying the determined textual representation to a trained artificial intelligence-based regression model. The method further includes, determining a best fit model and a unigram preparation data value for the n-gram level based on the determined model score, an n-gram preparation data value, and a type of n-gram using a trained classification model, wherein the best fit model is one of pretrained existing models and a dynamically trained n-gram model, predicting a bias score and an accuracy score for the input dataset by applying the determined textual representation and the input dataset onto the determined best fit model using a trained clustering model, generating a Prompt Inclusivity Index (PII) value based on the determined best fit model, the predicted bias score and the accuracy score, wherein the PII value indicates presence of bias data in the received input dataset, and outputting the generated PII value to a user via a user interface of a user device.