Bilal's team
To investigate Sapir-Whorf hypothesis by creating an LLM based on a logical constructed language.
Core trick
Trains a roughly 50k-parameter language model on a constructed Lojban corpus to test whether a logical grammar yields grammatical output.
Description
To investigate Sapir-Whorf hypothesis by creating an LLM based on a logical constructed language. Our findings show that even a very small model(50k params) is able to construct grammatically correct Lojban sentences.