HARP: Efficient Data Selection for Finetuning Large Language Models

2026-06-05 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Hierarchical Active Region Pruning (HARP) is an efficient, train-based data selection method designed for finetuning large language models. It addresses the challenge of balancing effective data selection for downstream objectives with the high cost of repeated model finetuning. Unlike scalable but proxy-reliant train-free selectors, or costly train-based methods requiring many train-evaluate iterations, HARP organizes the training data into a node-leaf hierarchy. It evaluates only representative leaves and infers unmeasured utilities using empirical Bayes posteriors. HARP then selects data via two envelopes: HARP-C for conservative redundancy control and HARP-E for additive complementary region rewards. Theoretically, HARP controls selection error and reduces train-evaluate costs under local smoothness and bounded estimation error. HARP variants outperform strong baselines by up to +8.9 points, utilizing approximately 7x fewer training examples.

Key takeaway

For Machine Learning Engineers optimizing large language model finetuning, HARP offers a significant efficiency gain. You can achieve superior downstream performance, up to +8.9 points, while drastically reducing training data requirements by approximately 7x. Consider integrating HARP's hierarchical data selection and empirical Bayes utility inference to streamline your finetuning workflows and control computational expenses. This approach allows for more effective data curation without extensive train-evaluate cycles.

Key insights

HARP efficiently selects finetuning data for LLMs by hierarchically evaluating representative subsets and inferring utilities.

Principles

• Balance data utility with selection cost.
• Hierarchical data organization reduces evaluation overhead.
• Empirical Bayes infers unmeasured data utility.

Method

HARP organizes data into a node-leaf hierarchy, evaluates representative leaves, infers unmeasured utilities with empirical Bayes, then selects data using HARP-C (redundancy control) or HARP-E (complementary rewards).

In practice

• Apply HARP to reduce LLM finetuning costs.
• Use HARP-C for redundancy-controlled data selection.
• Use HARP-E for complementary data selection.

Topics

• Large Language Models
• Finetuning
• Data Selection
• HARP
• Machine Learning Efficiency
• Empirical Bayes

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• CAPS: Cascaded Adaptive Pairwise Selection for Efficient Parallel Reasoning — CAPS optimizes LLM parallel reasoning by adaptively reducing verifier compute through cascaded pairwise selection.
• I Fine-Tuned DistilBERT Twice. It Humbled Me Both Times. — Data characteristics fundamentally determine fine-tuning performance and metric interpretation, not just the model.
• Demystifying Data Organization for Enhanced LLM Training — Effective data organization, beyond selection, significantly enhances LLM training stability and performance with minimal overhead.
• Efficient Test-Time Finetuning of LLMs via Convex Reconstruction and Gradient Caching — HullFT uses convex reconstruction and gradient caching to make test-time finetuning of LLMs faster and more efficient.
• Unified Data Selection for LLM Reasoning — High-Entropy Sum (HES) is a training-free metric for quantifying LLM reasoning quality via token entropy.
• Mining Useful General Data for Low-Resource Domain Adaptation — Neural Tangent Kernels can effectively select high-value auxiliary data for low-resource LLM domain adaptation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.