Fylo›ARCADIA›Graph
Hubs
InferenceChain·arcadia

Data leakage and training data bias impact model performance

This inference chain explains how naive split strategies, sequence filtering stringency, and validation set similarity modulate data leakage, which boosts model performance, and how biases in training data structure and curation couple with species abundance to generate performance biases in protein language models.

Confidence
90%
◑partialactivecomplexity: mid

Reasoning Steps (2)

Sequence similarity filtering reduces data leakageStep 1
Training data structure and curation biases influence model performance biasesStep 2

Source

Synthesis for current paper

Connections (6)

Naive split increases data leakageAssociation
Filtering stringency affects data leakageAssociation
Higher validation set similarity increases data leakageAssociation
Training data structure influences performance biasAssociation
User-level data curation bias affects performanceAssociation
Species abundance causes model biasAssociation