Task-Specific LLM Training With Zero-Shot Active Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Fine-tuning large language models (LLMs) for Natural Language Processing (NLP) tasks requires thousands or millions of annotations, leading to high costs.
Innovation Solution
A learning method that initializes with zero-shot learning using an unlabeled dataset to derive a labeled set, followed by an active learning loop to train a task-specific LLM, utilizing off-the-shelf LLMs to reduce annotation costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If fine-tuning LLMs on NLP tasks is performed, then human-level performance is achieved, but annotation costs increase significantly
Solution Approach 1:
The patent applies preliminary action by using zero-shot learning to pre-annotate data before the main fine-tuning process. The system performs initial predictions on unlabeled data using zero-shot LLMs, identifies high-confidence predictions, and uses these pre-annotated data to seed the active learning process, thereby reducing the total annotation cost while maintaining performance
Solution Approach 2:
The patent implements self-service by enabling the system to automatically generate and refine its own training data through active learning. The system queries an oracle to obtain labels for uncertain predictions, uses these labels to train task-specific models, and iteratively improves performance without requiring manual annotation of all training data, thus reducing external annotation costs
2Reliability
If thousands or millions of annotations are used for fine-tuning, then task-specific performance improves, but the cost of adapting LLMs increases
Solution Approach 1:
The patent extracts only the necessary annotations needed for effective fine-tuning by using active learning to identify and query only the most informative data points. Instead of annotating all available data, the system extracts and annotates only the high-uncertainty predictions that provide maximum learning benefit, significantly reducing the total number of annotations required while maintaining task-specific performance
Solution Approach 2:
The patent introduces an intermediary oracle mechanism that mediates between the system's predictions and the final training labels. The oracle provides ground truth labels only for the subset of data points where the system is uncertain, enabling efficient training with minimal annotations while achieving task-specific performance
Data Source
AI summary
The present disclosure provides a learning method, which includes steps as follows. An initialization is performed to use at least one large language model (LLM) with a zero-shot learning through an unlabeled dataset, so as to derive a labeled set. An active learning loop is performed to train a task-specific LLM (TLLM) through labeled set.