Minibatch Selection via Partition Matroid Constrained Gradient Matching

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

Summary

A new minibatch selection method, PartitionSel, is introduced for fine-tuning Large Language Models (LLMs) on heterogeneous datasets. This approach addresses the challenge of balancing rapid convergence with comprehensive data domain coverage, a limitation of existing independent or computationally expensive proxy-model methods. PartitionSel maximizes a validation-guided gradient-matching utility, incorporating per-domain budgets through a partition-matroid constraint. This design couples selections across domains, aiming to reduce redundancy and ensure more compatible training updates. The method's objective is weakly submodular, enabling an orthogonal matching pursuit algorithm with provable approximation guarantees. Empirical evaluations demonstrated PartitionSel's robust gains over per-domain and domain-agnostic baselines when fine-tuning Qwen2.5 and Llama-3 on MetaMathQA and Mol-Instructions benchmarks. It also effectively reduced conflicting gradient pairs within batches.

Key takeaway

For Machine Learning Engineers fine-tuning Large Language Models on heterogeneous data, consider implementing PartitionSel to optimize minibatch selection. This method demonstrably improves convergence and domain coverage compared to simpler baselines, reducing conflicting gradient pairs within batches. Your training updates will be more compatible, leading to robust performance gains on benchmarks like MetaMathQA and Mol-Instructions. Evaluate its integration to enhance your LLM fine-tuning workflows.

Key insights

PartitionSel optimizes LLM fine-tuning on heterogeneous data by balancing domain coverage and convergence via gradient-matching and matroid constraints.

Principles

• Gradient matching improves minibatch selection.
• Partition matroids constrain per-domain budgets.
• Cross-domain coupling reduces selection redundancy.

Method

PartitionSel maximizes a validation-guided gradient-matching utility under per-domain budgets, encoded as a partition-matroid constraint. It uses an orthogonal matching pursuit algorithm for optimization, ensuring cross-domain coupling and reduced redundancy.

In practice

• Fine-tune LLMs on diverse datasets.
• Improve training update compatibility.
• Outperform domain-agnostic baselines.

Topics

• Minibatch Selection
• Large Language Models
• Gradient Matching
• Partition Matroids
• LLM Fine-tuning
• Heterogeneous Data

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

Related on AIssential

• 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.
• 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.
• Learning What Matters: Probabilistic Task Selection via Mutual Information for Model Finetuning — TaskPGM optimizes LLM fine-tuning data mixtures by balancing task representativeness and diversity through an energy-based probabilistic framework.
• Beyond Cross-Validation: Adaptive Parameter Selection for Kernel-Based Gradient Descents — This paper proposes a novel Hybrid Selection Strategy (HSS) for Kernel-Based Gradient Descent (KGD) algorithms, which integrates bias-variance analysis with the splitting method for adaptive parameter selection.
• DualOptim+: Bridging Shared and Decoupled Optimizer States for Better Machine Unlearning in Large Language Models — DualOptim+ adaptively balances shared and decoupled optimizer states for superior LLM unlearning.
• GRZO: Group-Relative Zeroth-Order Optimization for Large Language Model Fine-Tuning — GRZO significantly reduces gradient estimation variance in zeroth-order optimization for memory-efficient LLM fine-tuning.

Open in AIssential →

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