Instruction Similarity Filtering for Zero-Shot AI Training Data

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for improving zero-shot learning performance in AI models through instruction tuning require excessive computing resources and may lead to negative transfer due to irrelevant training tasks, affecting model performance.

Innovation Solution

A system and method for constructing a dataset that evaluates similarity between training and target task instructions using cosine similarity or model transfer, selecting instructions with high similarity, and optionally using a pre-tuned task selector model to enhance accuracy, while removing placeholders to prevent negative effects.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If all instructions and instances from tasks are used for training to improve zero-shot learning performance, then model performance is improved, but computing resources are excessively consumed

Engineering Contradiction:
Improvezero-shot learning performanceVSAvoidcomputing resources
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent extracts only the instruction portion from task data, selectively taking out the relevant component (instructions) while excluding less useful components (instances). This extraction enables the model to learn task structures and perform zero-shot learning without processing all task data, thereby reducing computing resource consumption while maintaining performance improvement.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent segments task data into distinct components (instructions and instances) and processes only the instruction segment for training. This segmentation allows the model to focus on learning task definitions and structures from instructions alone, avoiding the computational burden of processing all task instances while still achieving zero-shot learning capabilities.

Inventive Principle:
Principle #1Segmentation

2Quantity of substance

If training tasks with no relevance to target task are included to increase training data volume, then training data quantity is increased, but negative transfer occurs degrading model performance

Engineering Contradiction:
Improvetraining data quantityVSAvoidmodel performance
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent applies local quality by selecting and training only on instructions that are locally relevant to the target task. Instead of uniformly processing all task instructions, the system identifies and focuses on instructions with high relevance to the target task, thereby increasing training data quantity for relevant tasks while avoiding the inclusion of irrelevant tasks that would cause negative transfer.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4682771A1System, method, and program for constructing data set for ai model training through instruction tuning
Publication Date: 2026.01.21 LG MANAGEMENT DEV INST CO LTD
  • EP4682771A1 patent drawingFigure 1
  • EP4682771A1 patent drawingFigure 2~3
  • EP4682771A1 patent drawingFigure 4

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 are disclosed. The method includes extracting 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, evaluating similarity by comparing the extracted instruction of the training task with the extracted instruction of the target task, selecting, from among the extracted instructions of the training task, instructions having a similarities equal to or greater than a predetermined value, and outputting the selected instructions of the training task as a data set.