larql

LARQL is a tool that decompiles transformer models into a queryable format known as a vindex, enabling users to query neural network weights as if they were working with a graph database. It leverages Lazarus Query Language (LQL) to browse, edit, and recombine model knowledge without the need for GPUs, and offers various extraction options for inference and browsing, alongside features like HTTP and gRPC serving capabilities. Notable functionalities include the ability to describe or extract model features and infer relationships directly through conversational queries.