US 2026/0187112 A1Published application

Device and system for editing trained ai models

AssigneeMultiverse Computing S.L.

InventorsRoman ORUS, Rodrigo CIFUENTES HERNÁNDEZ

Life of this patent

Estimating the term…

Loading patent…

Abstract

A device for editing large language models and multimodal AI models. The device includes a memory medium that stores the data and/or dataset of the model, a control unit that is configured to process the data within modules and according to requests and pre-defined configuration parameters from a user and/or client which communicates with the device with at least one communication unit to provide at least one output based on an input data; and runs the instruction on the trained AI model. The device includes a compression module, a lobotomization module and an inception module and at least one of the modules is configured to identify layers of the trained model with the weight matrices; and decomposing the weight matrices of the trained model into a tensor network structure. The device also includes an integration module that provides connection and data transfer between external systems and platforms and the device.

Claims14

Dependent claims

A device for editing large language models and multimodal AI models, the device comprising:

at least one memory medium that stores the data and/or dataset of the model;
at least one control unit configured to process the data within modules and according to requests and pre-defined configuration parameters from at least one user and/or client which communicates with the device with at least one communication unit to provide at least one output based on an input data;
and runs the instruction on the trained AI model;
at least one compression module;
at least one lobotomization module;
at least one inception module, wherein at least one of the at least one compression module, the at least one lobotomization module and the at least one inception module is configured to identify layers of the trained model with the weight matrices and decompose the weight matrices of the trained model into a tensor network structure, wherein the compression module is configured to compress the size of the pre-trained layers of the tensorized model with compressing the decomposed weight matrices and to store the tensor network structure in the memory medium, wherein the lobotomization module is configured to inject information to the trained model with encoding data values of the dataset in one or more tensors of the trained model associated with values of the dataset, and wherein the at least one inception module is configured to remove information to the trained model with encoding zero data values of the dataset or removing the one or more tensor associated with values of the dataset to provide removing data that defined by user and/or client from the trained model;
at least one integrity detection module that has a baseline comparison and hashing unit configured to hash the model and re-hash the model periodically, and comparing the re-hashed value and first hashed value to tracking the modification status of the trained model;
and
at least one integration module that provides connection and data transfer between external systems and platforms and the device.

6 more elements

A system for editing large language models and multimodal AI models, the system including:

at least one computer including at least one memory medium that stores the data and/or dataset of the model;
at least one control unit configured to process the data within modules and according to requests and pre-defined configuration parameters from at least one user and/or client which communicates with the system with at least one communication unit to provide at least one output based on an input data and also establishing communication between the computers within the system;
at least one compression module;
at least one lobotomization module;
at least one inception module, wherein at least one of the at least one compression module, the at least one lobotomization module and the at least one inception module is configured to identifying layers of the trained model with the weight matrices;
and decomposing the weight matrices of the trained model into a tensor network structure, wherein the compression module configured to compress the size of the pre-trained layers of the tensorized model with compressing the decomposed weight matrices;
and
storing the tensor network structure in the memory medium, wherein the lobotomization module is configured to inject information to the trained model with encoding data values of the dataset in one or more tensors of the trained model associated with values of the dataset, and wherein the at least one inception module is configured to remove information to the trained model with encoding zero data values of the dataset or removing the one or more tensor associated with values of the dataset to provide removing data that defined by user and/or client from the trained model;
at least one integrity detection module that has a baseline comparison and hashing unit that is configured to hash the model and re-hash the model periodically, and compare the re-hashed value and first hashed value to tracking the modification status of the trained model;
and
at least one integration module that provides connection and data transfer between external systems and platforms and the device.

8 more elements

Dependencies are inferred from the claim text.