CLCVLGNov 29, 2024

On Domain-Adaptive Post-Training for Multimodal Large Language Models

arXiv:2411.19930v413 citationsh-index: 21Has CodeEMNLP
Originality Incremental advance
AI Analysis

This addresses the problem of domain adaptation for MLLMs in fields like biomedicine and remote sensing, offering an incremental improvement over existing adaptation techniques.

The paper tackles adapting multimodal large language models to specific domains via post-training, finding that a single-stage training approach and a generate-then-filter data synthesis method outperform existing methods, with open-sourced models, code, and data.

Adapting general multimodal large language models (MLLMs) to specific domains, such as scientific and industrial fields, is highly significant in promoting their practical applications. This paper systematically investigates domain adaptation of MLLMs via post-training, focusing on data synthesis, training pipeline, and task evaluation. (1) Data Synthesis: Using only open-source models, we develop a generate-then-filter pipeline that curates diverse visual instruction tasks based on domain-specific image-caption pairs. The resulting data surpass the data synthesized by manual rules or strong closed-source models in enhancing domain-specific performance. (2) Training Pipeline: Unlike general MLLMs that typically adopt a two-stage training paradigm, we find that a single-stage approach is more effective for domain adaptation. (3) Task Evaluation: We conduct extensive experiments in high-impact domains such as biomedicine, food, and remote sensing, by post-training a variety of MLLMs and then evaluating MLLM performance on various domain-specific tasks. Finally, we fully open-source our models, code, and data to encourage future research in this area.

Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes