Learning self-evaluation to improve selective prediction in LLMs
Aspects of the disclosure are directed to methods, systems, and computer readable media for adaptation with self-evaluation to improve selective prediction in large language models (LLMs), generally referred to as ASPIRE. ASPIRE includes training LLMs on a portion of training data from a question answering task to learn self-evaluation, e.g., learn to distinguish whether a generated answer is correct or not. ASPIRE further includes a selection score that combines a likelihood of that generated answer is correct with a self-evaluation score for selective prediction. ASPIRE demonstrates improved selective prediction performance with less computational cost.
1 . A method for selective prediction, comprising:
training, by one or more processors, a large language model (LLM) to a task by adjusting first adaptable parameters to the task using training data;
generating, by one or more processors, a plurality of outputs associated with the task using the LLM with the adjusted first adaptable parameters;
training, by the one or more processors, the LLM on self-evaluation by:
freezing the adjusted first adaptable parameters;
determining whether each of the plurality of outputs is correct using an evaluation metric; and
adjusting second adaptable parameters for the LLM based on the determination; and
generating, by the one or more processors, a prediction for the task using the LLM based on the adjusted first adaptable parameters and adjusted second adaptable parameters, the prediction comprising a self-evaluation score.
2 . The method of claim 1 , wherein the LLM is a pretrained LLM.
3 . The method claim 1 , wherein training the LLM to the task and on self-evaluation comprises fine-tuning the LLM using soft prompt tuning.
4 . The method of claim 1 , wherein training the LLM to the task further comprises:
freezing model parameters of the LLM; and
adding and iteratively updating the first adaptable parameters for the LLM.
5 . The method of claim 1 , wherein training the LLM on self-evaluation further comprises:
freezing model parameters of the LLM; and
adding and iteratively updating the second adaptable parameters for the LLM.
6 . The method of claim 5 , wherein determining whether each of the plurality of outputs is correct further comprises labeling each of the plurality of outputs as correct or wrong.
7 . The method of claim 5 , wherein the evaluation metric compares a similarity of an output of the plurality of outputs to a reference output.
8 . The method of claim 7 , wherein determining whether each of the plurality of outputs is correct further comprises determining whether the output is within a threshold of the reference output.
9 . The method of claim 8 , wherein the threshold is a Rouge threshold having a value large enough where outputs that are wrong are not determined to be correct.
10 . A system comprising:
one or more processors; and
one or more storage devices coupled to the one or more processors and storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations for selective prediction, the operations comprising:
training a large language model (LLM) to a task by adjusting first adaptable parameters to the task using training data;
generating a plurality of outputs associated with the task using the LLM with the adjusted first adaptable parameters;
training the LLM on self-evaluation by:
freezing the adjusted first adaptable parameters;
determining whether each of the plurality of outputs is correct using an evaluation metric; and
adjusting second adaptable parameters for the LLM based on the determination; and
generating a prediction for the task using the LLM based on the adjusted first adaptable parameters and adjusted second adaptable parameters, the prediction comprising a self-evaluation score.
11 . The system of claim 10 , wherein the LLM is a pretrained LLM.
12 . The system of claim 10 , wherein training the LLM to the task and on self-evaluation comprises fine-tuning the LLM using soft prompt tuning.
13 . The system of claim 10 , wherein training the LLM to the task further comprises:
freezing model parameters of the LLM; and
adding and iteratively updating the first adaptable parameters for the LLM.
14 . The system of claim 10 , wherein training the LLM on self-evaluation further comprises:
freezing model parameters of the LLM; and
adding and iteratively updating the second adaptable parameters for the LLM.
15 . The system of claim 14 , wherein determining whether each of the plurality of outputs is correct further comprises labeling each of the plurality of outputs as correct or wrong.
16 . The system of claim 14 , wherein the evaluation metric compares a similarity of an output of the plurality of outputs to a reference output.
17 . The system of claim 16 , wherein determining whether each of the plurality of outputs is correct further comprises determining whether the output is within a threshold of the reference output.
18 . The system of claim 17 , wherein the threshold is a Rouge threshold having a value large enough where outputs that are wrong are not determined to be correct.
19 . A non-transitory computer readable medium for storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations for selective prediction, the operations comprising:
training a large language model (LLM) to a task by adjusting first adaptable parameters to the task using training data;
generating a plurality of outputs associated with the task using the LLM with the adjusted first adaptable parameters;
training the LLM on self-evaluation by:
freezing the adjusted first adaptable parameters;
determining whether each of the plurality of outputs is correct using an evaluation metric; and
adjusting second adaptable parameters for the LLM based on the determination; and
generating a prediction for the task using the LLM based on the adjusted first adaptable parameters and adjusted second adaptable parameters, the prediction comprising a self-evaluation score.
20 . The non-transitory computer readable medium of claim 19 , wherein training the LLM to the task and on self-evaluation comprises fine-tuning the LLM using soft prompt tuning.