Executing a client model using a task prompt produced by a main system
A technique executes a client machine-trained model (“client model”) on a client device. In operation, the client device submits a description of a task to be performed by the client device to a network-accessible main system. The main system uses a main-system machine-trained model (“main-system model”) to produce a task prompt based on the task description. The client device subsequently uses the task prompt to process queries pertaining to the task. The main-system is trained to increase the accuracy of responses produced by the client model, while reducing the sizes of task prompts produced by the main system. The training process is performed by holding weights of the client model constant.
1 . A method for interacting with a client machine-trained model provided at a client device, comprising:
receiving a task description that describes a task to be performed using the client machine-trained model;
submitting the task description to a network-accessible main system, wherein the main system includes a main-system machine-trained model, wherein the main-system machine-trained model, as a result of a training process that was performed, includes weights that have learned a vector space associated with the client machine-trained model, wherein weights of the client machine-trained model were held constant throughout the training process that was performed,
wherein the main-system machine-trained model has been trained in the training process by: transforming task descriptions into task prompts associated with respective tasks; receiving model responses to queries produced by the client machine-trained model using the task prompts; determining measures of prediction loss by comparing the model responses to associated ground-truth responses, to generate measures of prediction loss; and using the measures of prediction loss to update the weights of the main-system machine-trained model without training the weights of the client machine-trained model in the training process;
receiving a task prompt from the main system, the main-system machine-trained model producing the task prompt based on the task description that is submitted and based on the learned vector space associated with the client machine-trained model, the task prompt being embedding information that expresses a meaning of the task description; and
using the task prompt to process queries associated with the task,
wherein the using includes, for each query of the queries associated with the task:
receiving the query;
producing a query prompt based on the query;
producing a combined prompt by combining the query prompt with the task prompt; and
using the client machine-trained model to produce a model response to the query based on the combined prompt.
2 . The method of claim 1 , wherein the client device is a user computing device or a local system that includes one or more client devices, wherein the main system is implemented by one or more servers, and wherein the using the task prompt occurs offline, without interaction with the main system.
3 . The method of claim 1 , wherein the task description that is received includes a description of an objective of the task, and plural examples of model responses that are to be generated upon the submission of respective queries, in accordance with the task.
4 . The method of claim 1 , further comprising acquiring a pre-generated task prompt from a network-accessible data store, and using the pre-generated task prompt to interact with the client machine-trained model.
5 . The method of claim 1 , wherein the main-system machine-trained model includes a base machine-trained model that provides an initial task prompt, and a prompt-compressing machine-trained model that compresses the initial task prompt, to produce the task prompt that is sent to the client device.
6 . The method of claim 1 ,
wherein the main-system machine-trained model has been trained in the training process based on a measure of loss that is a combination of the prediction loss and sparsity loss,
wherein the sparsity loss measures removal of redundant information from the task prompts.
7 . A main system for interacting with a client device, comprising:
a data store for storing computer-readable instructions;
a processing system for executing the computer-readable instructions in the data store, to perform operations including:
receiving a task description from the client device over a computer network, the task description describing a task to be performed by the client device using a client machine-trained model;
using a main-system machine-trained model to transform the task description into a task prompt, the task prompt being embedding information that expresses a meaning of the task description, wherein the main-system machine-trained model, as a result of a training process that was performed, includes weights that have learned a vector space associated with the client machine-trained model, wherein weights of the client machine-trained model were held constant throughout the training process that was performed,
wherein the main-system machine-trained model has been trained in the training process by: transforming task descriptions into task prompts associated with respective tasks; receiving model responses to queries produced by the client machine-trained model using the task prompts; determining measures of prediction loss by comparing the model responses to associated ground-truth responses, to generate measures of prediction loss; and using the measures of prediction loss to update the weights of the main-system machine-trained model without training the weights of the client machine-trained model in the training process;
sending the task prompt over the computer network to the client device, wherein the client machine-trained model uses the task prompt to transform queries pertaining to the task into model responses by transforming the queries into query prompts, producing combined prompts by combining the query prompts with the task prompt, and producing the model responses based on the combined prompts,
wherein the client machine-trained model has fewer weights than the main-system machine-trained model.
8 . The main system of claim 7 , wherein the main system is implemented by one or more servers, and the client device is a user computing device.
9 . The main system of claim 7 , wherein the task description that is received includes a description of an objective of the task, and plural examples of model responses that are to be generated upon the submission of respective queries, in accordance with the task.
10 . The main system of claim 7 , wherein the main-system machine-trained system includes:
a base machine-trained model that transforms the task description into an initial task prompt; and
a prompt-compressing machine-trained model compresses the initial task prompt, to produce the task prompt that is sent to the client device.
11 . The main system of claim 7 ,
wherein the main-system machine-trained model has been trained in the training process based on a measure of loss that is a combination of the prediction loss and sparsity loss,
wherein the sparsity loss measures removal of redundant information from the task prompt prompts.
12 . The main system of claim 11 ,
wherein measures of the sparsity loss are based on numbers of elements in the task prompts.
13 . A computer-readable storage medium for storing computer-readable instructions, a processing system executing the computer-readable instructions to perform operations in a client device, the operations comprising:
receiving a task description that describes a task to be performed using a client machine-trained model of the client device;
submitting the task description to a network-accessible main system, the main system providing a main-system machine-trained model;
receiving a task prompt from the main system, the main-system machine-trained model producing the task prompt based on the task description, the task prompt being embedding information that expresses a meaning of the task description,
wherein the main-system machine-trained model, as a result of a training process that was performed, includes weights that have learned a vector space associated with the client machine-trained model, wherein weights of the client machine-trained model were held constant throughout the training process that was performed,
wherein the main-system machine-trained model has been trained in the training process by: transforming task descriptions into task prompts associated with respective tasks; receiving model responses to queries produced by the client machine-trained model using the task prompts; determining measures of prediction loss by comparing the model responses to associated ground-truth responses, to generate measures of prediction loss; and using the measures of prediction loss to update the weights of the main-system machine-trained model without training the weights of the client machine-trained model in in the training process; and
using the task prompt to process queries associated with the task, wherein the using includes, for each query of the queries associated with the task:
receiving the query;
producing a query prompt based on the query;
producing a combined prompt by combining the query prompt with the task prompt; and
using the client machine-trained model to produce a model response to the query based on the combined prompt.
14 . The method of claim 1 , further comprising using the client machine-trained model to produce responses for another task for which the main-system machine-trained model has not been trained in the training process.
15 . The method of claim 1 , wherein each part of the task prompt is an embedding that represents a prompt token, wherein the query prompt is produced using a vocabulary of text-based tokens, and wherein each prompt token differs from each of the text-based tokens.
16 . The method of claim 5 , wherein the training process that was performed involves training weights of the prompt-compressing machine-trained model while holding weights of the base machine-trained model and the weights of the client machine-trained model fixed throughout the training process.
17 . The main system of claim 10 , wherein the training process that was performed involves training weights of the prompt-compressing machine-trained model while holding weights of the base machine-trained model and the weights of the client machine-trained model fixed throughout the training process.
18 . The method of claim 1 , wherein the weights of the main-system machine-trained model are capable of generating task prompts for additional task descriptions not included in the training process, without requiring additional training based on training examples pertaining to the additional task descriptions.