Learning from Generalization Patterns: An Evaluation-Driven Approach to Enhanced Data Augmentation for Fine-Tuning Small Language Models

2025-10-20

AI Summary

Small language models (SLMs) offer deployment efficiency advantages but suffer from poor generalization and low accuracy on complex domain tasks; moreover, supervised fine-tuning relies heavily on costly human annotations. To address this, we propose an evaluation-driven data augmentation method: leveraging validation-set failure cases to automatically identify structured generalization error patterns and generate targeted training samples, thereby enabling closed-loop data optimization. Our key contribution is the first systematic extraction of interpretable, structured failure patterns—not from training loss, but directly from model generalization errors—and the design of pattern-guided augmentation strategies grounded in these insights. The approach is implemented within the PaDA-Agent framework, which unifies evaluation, diagnostic analysis, and synthetic data generation modules. Experiments on Llama 3.2 1B Instruct demonstrate that our method significantly outperforms existing large-model–driven data augmentation techniques, effectively narrowing the generalization gap and enhancing fine-tuning performance.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search
• Natural Language Processing: (Large) Language Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Small Language Models (SLMs) offer compelling advantages in deployment cost and latency, but their accuracy often lags behind larger models, particularly for complex domain-specific tasks. While supervised fine-tuning can help bridge this performance gap, it requires substantial manual effort in data preparation and iterative optimization. We present PaDA-Agent (Pattern-guided Data Augmentation Agent), an evaluation-driven approach that streamlines the data augmentation process for SLMs through coordinated operations. Unlike state-of-the-art approaches that focus on model training errors only and generating error-correcting samples, PaDA-Agent discovers failure patterns from the validation data via evaluations and drafts targeted data augmentation strategies aiming to directly reduce the generalization gap. Our experimental results demonstrate significant improvements over state-of-the-art LLM-based data augmentation approaches for Llama 3.2 1B Instruct model fine-tuning.

Problem

• Enhancing data augmentation for small language models
• Reducing accuracy gap between small and large models
• Automating pattern-driven data augmentation strategies

Innovation

• Evaluation-driven data augmentation for small language models
• Pattern-guided agent discovers failure patterns from validation
• Targeted strategies directly reduce generalization gap