Technical Manual for a Toolkit for Measuring Contextual Individuation in Transformer Language Models
Addressing the challenge of verifying how transformer models separate word senses across contexts, this open toolkit introduces 'bridge forms'—identical word tokens spanning disparate domains—as controlled test constructs. The technical manual details an end-to-end pipeline spanning Wikipedia corpus extraction, layer-wise token localization, pairwise silhouette separation metrics, and visualization protocols, while outlining safeguards against subword misalignment and dimensionality reduction artifacts. It serves strictly as a measurement instrument without presenting empirical model evaluations.
Why it's worth reading
It provides a rigorous, artifact-controlled methodology and open pipeline for researchers probing how language models represent contextual word senses across layers.