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

Researchers propose a framework for optimizing data selection in large language model instruction tuning by learning task-specific and model-specific weights for multiple quality indicators. Using efficient in-context learning signals on small validation sets, the method achieves comparable performance to full-dataset training with only 30% of samples, revealing important trade-offs between semantic diversity and logical complexity.

This research addresses a fundamental efficiency challenge in large language model development: the computational cost and resource requirements of instruction tuning. Rather than treating all training data equally or applying static weighting schemes, the proposed framework adapts data selection dynamically based on both the target downstream task and the specific model architecture. This represents a meaningful shift from one-size-fits-all approaches to personalized optimization strategies.

The breakthrough lies in using in-context learning signals on compact validation sets as performance proxies. This allows researchers to identify optimal weighting configurations without expensive full-scale fine-tuning cycles, dramatically reducing computational overhead. Validation across multiple model families—Mistral, Qwen, and Llama—demonstrates the approach's generalizability, achieving 30% sample efficiency on GSM8K benchmarks without sacrificing performance.

For AI developers and organizations, this directly impacts operational costs and resource allocation. Reducing training data requirements from 100% to 30% translates to faster iteration cycles, lower GPU utilization, and accelerated model deployment. The research reveals nuanced insights about reasoning tasks, showing that semantic diversity and logical complexity operate in tension—an observation that challenges assumptions about what makes effective training data.

The implications extend beyond efficiency metrics. As models become larger and more capable, the ability to precisely target which data matters most creates competitive advantages in both commercial and research contexts. This framework could influence how organizations structure their data pipelines and quality assessment processes moving forward.

• →A framework learns task-specific and model-specific weights for data selection, replacing static weighting schemes with adaptive optimization.
• →In-context learning signals on tiny validation sets serve as efficient performance proxies, eliminating the need for expensive full-scale fine-tuning.
• →The method achieves equivalent performance to full-dataset training using only 30% of samples on GSM8K benchmarks.
• →Testing across Mistral, Qwen, and Llama models confirms the approach generalizes across different architectures.
• →Analysis reveals a trade-off between semantic diversity and logical complexity in reasoning tasks, informing future data curation strategies.