Training Data Bias
A systematic bias in machine learning models arising from uneven distribution of training data, causing models to produce stereotyped or templated outputs for underrepresented categories. It is considered a core technical problem affecting the diversity of AI-generated content.
Core Facts
Timeline (last 90 days)
Training data bias refers to a mismatch between the statistical associations a model learns and real-world causality.
If articles from a certain type of media outlet frequently appear in the context of "how to solve a certain technical problem" in the corpus, the model may bind that media outlet to technical support, forming a spurious association.
If the AI model has training data bias, such as being friendlier to a certain expression style or accent, it may inadvertently screen out candidates who should have passed.
Methods for overfitting benchmarks include mixing evaluation data into training corpora, indirect web-crawl contamination, instruction-style fine-tuning, and training models to recognize evaluation scenarios.
All Facts (4)
Methods for overfitting benchmarks include mixing evaluation data into training corpora, indirect web-crawl contamination, instruction-style fine-tuning, and training models to recognize evaluation scenarios.
65%UnverifiedTraining data bias refers to a mismatch between the statistical associations a model learns and real-world causality.
50%UnverifiedIf articles from a certain type of media outlet frequently appear in the context of "how to solve a certain technical problem" in the corpus, the model may bind that media outlet to technical support, forming a spurious association.
50%UnverifiedIf the AI model has training data bias, such as being friendlier to a certain expression style or accent, it may inadvertently screen out candidates who should have passed.
50%