Fine-tuned artificial intelligence model with event mapping features

1 . A method, comprising:

obtaining a first conversation stream of a first conversation, wherein the first conversation stream includes one or more messages and one or more events;

mapping an event of the one or more events to at least one message of the one or more messages, wherein the mapping is performed using event-based labeling and comprises:

determining a first message of the at least one message that is closest in time to the event and precedes the event;

determining one or more additional messages of the at least one message that precede the first message, wherein a time delta between an earliest message of the one or more additional messages and the first message is within a threshold; and

determining an event type corresponding to the event; and

fine-tuning a language model (LM) according to the mapping of the event to the at least one message at least by using the event type as a label that is input to the LM and the at least one message as input data to the LM, the fine-tuning of the LM comprising adjusting one or more LM parameters such that the fine-tuned LM is configured for:

determining one or more tools that are external to the fine-tuned LM and invokable, by the fine-tuned LM, at least to generate an output that is related to the event type, the event, and the at least one message, the output being usable by the fine-tuned LM during a second conversation to perform one or more actions.

2 . The method of claim 1 , wherein the event type is a ground truth label for determining the one or more tools.

3 . The method of claim 1 , wherein:

the first message and the one or more additional messages comprise transcript data, and

the mapping further comprises joining the transcript data with the event.

4 . The method of claim 1 , wherein the event is another label usable to fine-tune the LM.

5 . The method of claim 1 , wherein:

the mapping is performed further using content-based labeling, and

the mapping comprises:

determining a content-based label based on a similarity measure between transcript data of the at least one message and a selected message; and

selecting the content-based label as a tool of the one or more tools if the similarity measure exceeds a similarity threshold.

6 . The method of claim 1 , wherein fine-tuning the LM comprises setting an input limit and an output limit for a prompt of the LM based on a quantity of tokens, the input limit and the output limit being set to satisfy a latency threshold of the LM.

7 . The method of claim 6 , wherein the quantity of tokens is at least one of a maximum quantity of tokens that meets a generation criterion or a set of tokens associated with the one or more LM parameters.

8 . The method of claim 1 , further comprising:

deploying the fine-tuned LM in a system; and

using the deployed fine-tuned LM to:

determine the one or more tools while the second conversation is ongoing; and

perform the one or more actions using the one or more tools.

9 . The method of claim 8 , wherein the one or more actions are performed within an interval of time measured from a start time of a corresponding message of the second conversation.

10 . An apparatus comprising a processing system that includes one or more memories and one or more processors coupled to the one or more memories, the processing system configured to cause the apparatus to:

obtain a first conversation stream of a first conversation, wherein the first conversation stream includes one or more messages and one or more events;

map an event of the one or more events to at least one message of the one or more messages, wherein the mapping is performed using event-based labeling and the processing system that is configured to cause the apparatus to map the event of the one or more events to the at least one message of the one or more messages is further configured to:

determine a first message of the at least one message that is closest in time to the event and precedes the event;

determine one or more additional messages of the at least one message that precede the first message, wherein a time delta between an earliest message of the one or more additional messages and the first message is within a threshold; and

determine an event type corresponding to the event; and

fine-tune a language model (LM) according to the mapping of the event to the at least one message at least by using the event type as a label that is input to the LM and the at least one message as input data to the LM, wherein the processing system that is configured to cause the apparatus to fine-tune the LM is further configured to adjust one or more LM parameters such that the fine-tuned LM is configured to:

determine one or more tools that are external to the fine-tuned LM and invokable, by the fine-tuned LM, at least to generate an output that is related to the event type, the event, and the at least one message, the output being usable by the fine-tuned LM during a second conversation to perform one or more actions.

11 . The apparatus of claim 10 , wherein the event type is a ground truth label for determining the one or more tools.

12 . The apparatus of claim 10 , wherein the event is another label usable to fine-tune the LM.

13 . The apparatus of claim 10 , wherein the processing system that is configured to cause the apparatus to fine-tune the LM is further configured to set an input limit and an output limit for a prompt of the LM based on a quantity of tokens.