Instruction-Tuned Training Data Selection for Zero-Shot AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI models face challenges in zero-shot learning performance due to training tasks that are not relevant to the target task, leading to negative transfer and degraded performance.
Innovation Solution
A system and method for constructing a data set that improves zero-shot learning by extracting and selecting instructions from training tasks with similarities equal to or greater than a predetermined value, using cosine similarity and model transfer methods, and unifying placeholders into specific terms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If training tasks with no relevance to the target task are used, then the quantity of training data increases, but the zero-shot learning performance degrades due to negative transfer
Solution Approach 1:
The patent extracts and selects only the relevant training tasks that have similarity to the target task, removing irrelevant tasks that cause negative transfer. This is achieved through evaluating similarity between training tasks and target tasks, and selecting only those that meet a predetermined similarity threshold, thereby maintaining data quantity while improving performance.
Solution Approach 2:
The patent applies local quality by making different parts of the training data have different functions - irrelevant tasks are excluded while relevant tasks are retained. The selection process ensures that each included training task has specific relevance to the target task, creating a heterogeneous but purposeful training set that avoids negative transfer.
2Adaptability or versatility
If all training tasks are used without selection, then the diversity of training data increases, but the relevance to target task decreases leading to negative transfer
Solution Approach 1:
The patent implements feedback through the similarity evaluation mechanism that compares each training task with the target task. This feedback loop identifies which tasks are relevant and which are not, allowing the system to selectively include only those tasks that maintain both diversity and relevance to the target task.
Solution Approach 2:
The patent changes the parameter of task selection from inclusive (all tasks) to selective (tasks above similarity threshold). By adjusting the similarity threshold parameter, the system can control the balance between diversity and relevance, ensuring that only tasks meeting the relevance criterion are included in the training set.
3Adaptability or versatility
If placeholders are not unified into specific terms, then the flexibility of instruction templates is maintained, but the training effectiveness decreases due to negative impact of placeholders
Solution Approach 1:
The patent applies preliminary action by unifying placeholders into specific terms before the training process begins. This pre-processing step converts abstract placeholder references into concrete, meaningful terms, ensuring that the training data is optimized for effectiveness while the actual instruction templates remain flexible for different applications.
Data Source
AI summary
A system, method, and program for constructing a data set that improves zero-shot learning performance of an AI model through instruction tuning extract instructions from each of a training task used for training an AI model, and a target task that is a task to be trained through the training task, evaluate similarity by comparing the extracted instruction of the training task with the extracted instruction of the target task, select, from among the extracted instructions of the training task, instructions having a similarities equal to or greater than a predetermined value, and output the selected instructions of the training task as a data set.