Technology Category
Machine Learning: Large Multimodal Models (LMMs)Search and Optimization: Sampling/Simulation-based SearchNatural Language Processing: Code Generation / Program Synthesis from Natural Language
Stratified Selective Sampling for Instruction Tuning with Dedicated Scoring Strategy
🤖 AI Summary
To address inefficiency, poor generalization, and domain limitations in instruction tuning of large language models (LLMs), this paper proposes an efficient and generalizable hierarchical data selection framework. The method comprises a three-stage pipeline: (1) task-driven grouping followed by embedding-based clustering to ensure comprehensive task coverage and sample diversity; (2) lightweight, task-specialized quality assessment models to identify high signal-to-noise-ratio samples; and (3) difficulty-aware controllable sampling to jointly regulate dataset scale and difficulty distribution. We introduce the novel “hierarchical–grouping–evaluation–regulation” paradigm. Empirical results demonstrate that our approach reduces computational overhead for data selection by an order of magnitude while preserving—or even improving—multi-task generalization performance. Moreover, it enables the construction of high-quality, compact, and structurally controllable cross-domain instruction datasets.
Technology Category
Application Category
📝 Abstract
Recent work shows that post-training datasets for LLMs can be substantially downsampled without noticeably deteriorating performance. However, data selection often incurs high computational costs or is limited to narrow domains. In this paper, we demonstrate that data selection can be both -- efficient and universal -- by using a multi-step pipeline in which we efficiently bin data points into groups, estimate quality using specialized models, and score difficulty with a robust, lightweight method. Task-based categorization allows us to control the composition of our final data -- crucial for finetuning multi-purpose models. To guarantee diversity, we improve upon previous work using embedding models and a clustering algorithm. This integrated strategy enables high-performance fine-tuning with minimal overhead.
Problem
Research questions and friction points this paper is trying to address.
Efficient universal data selection for LLM instruction tuning
Balancing data quality and diversity with specialized scoring
Minimizing computational overhead in post-training dataset downsampling
Innovation
Methods, ideas, or system contributions that make the work stand out.
Multi-step pipeline for efficient data selection
Specialized models for quality estimation
Embedding models and clustering for diversity
🔎 Similar Papers
No similar papers found.
