Learning Multi-Indicator Weights for Data Selection: A Joint Task-Model Adaptation Framework with Efficient Proxies

2026-05-10

AI Summary

Existing data selection methods typically rely on static, task- and model-agnostic weights, which struggle to accommodate variations across downstream tasks and model architectures. This work proposes a task–model co-adaptive data selection framework that jointly optimizes both dimensions during data filtering—a first in the field. By leveraging in-context learning (ICL) signals from a small validation set, the approach constructs a lightweight proxy evaluator that dynamically combines multiple quality metrics, such as semantic diversity and logical complexity, with optimal adaptive weights. Experiments demonstrate that the method achieves comparable or superior performance to full-data fine-tuning on benchmarks like GSM8K using only 30% of the training data, substantially improving data efficiency.

Abstract

Data selection is a key component of efficient instruction tuning for large language models, as recent work has shown that data quality often matters more than data quantity. Accordingly, prior studies have introduced various multi-dimensional heuristics to evaluate and filter instruction data. However, most existing methods rely on static task-agnostic and model-agnostic weighting schemes, which overlook the varying requirements of specific downstream tasks and the differing pre-existing capabilities of models. In this paper, we propose a framework for learning multi-indicator weights that jointly adapts data selection to both the downstream task and the specific model. Our method identifies optimal weight configurations without full-scale fine-tuning by utilizing in-context learning (ICL) signals on compact tiny-validation sets. These signals serve as efficient performance proxies that ensure high-fidelity evaluation at minimal computational cost. Experiments across multiple benchmarks and model families, including Mistral, Qwen, and Llama, show that the approach achieves performance comparable to or exceeding full-dataset tuning while using only 30\% of the training samples on GSM8K. Furthermore, our analysis reveals a trade-off between semantic diversity and logical complexity in reasoning tasks, highlighting the necessity of joint task-model adaptation.

Problem

• data selection
• multi-indicator weighting
• task-model adaptation
• instruction tuning
• large language models

Innovation

• multi-indicator weighting
• task-model adaptation
• in-context learning (ICL)
• data selection
• efficient proxy evaluation