Hidden states guide data selection to tune large language models efficiently
Hidden Cartographer: Hidden State Guided Data Selection for Dual Budgeted Instruction Tuning
Computation and Language
Summary
Training large language models for specialized areas like medicine or law usually needs many expert-labeled examples, which is expensive. The authors propose a new method called Hidden Cartographer that picks the most useful data by looking inside the model’s hidden layers and understanding training effects precisely. This way, fewer examples need expert labeling and less data is needed for fine-tuning, saving time and money. Their tests show this method works better than others across several expert domains.
What this means in practice
- •For medical ai developers: Select the most impactful medical text data for annotation to efficiently fine-tune models for clinical tasks.
- •For legal tech teams: Cut annotation costs by precisely identifying key legal documents that improve model adaptation with limited expert input.
Authors
Zixuan Qin, Qi Qin, Erbo Li, Ting Wei, Yifan Sun, Zhaoxin Fan
Abstract
Given the high cost of expert annotation and limited fine-tuning resources, effective data selection from large unlabeled pools under strict budgets has become a crucial step in adapting LLMs to specialized domains. Existing label-dependent methods risk distorted selection when using pseudo-labels, while traditional active learning methods do not explicitly characterize instruction-tuning dynamics. To address these limitations, we propose Hidden Cartographer (HC), a two-stage framework that uses hidden states to guide annotation and exact training dynamics to select the final training subset. The first stage is Progressive Candidate Expansion, which utilizes the proxy to iteratively select top unlabeled candidates for expert annotation, merging them to recompute training dynamics and update the proxy until the annotation budget is exhausted; the second stage is Exact Cartography Pruning, which ranks these annotated candidates by their exact dynamics scores to isolate a compact subset under the final training budget. Experiments across the medical, legal, and financial domains demonstrate that HC achieves the highest average accuracy among all budgeted baselines, requiring annotation for only 20--30% of the original pool and fine-tuning on just 10--15%.