Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Huan Song, Deeksha Razdan, Yiyue Qian, Arijit Ghosh Chowdhury, Parth Patwa, Aman Chadha, Shinan Zhang, Sharlina Keshava, Hannah Marlowe

TL;DR

We address generalization gaps in small language models by introducing PaDA-Agent, an evaluation-driven data augmentation framework that links validation-time failures to targeted synthetic data generation. A central orchestrator coordinates Pattern Analysis, Data Generation, and Quality Control to discover generalization patterns, draft augmentation strategies, and ensure data utility. Empirical results on a 1B Instruct Llama model show consistent improvements across reasoning, knowledge, and coding tasks, with pronounced gains in low-data regimes and interpretable strategies that reveal underlying failure modes. The approach reduces manual data engineering by closing the generalization gap through a principled, evaluation-informed augmentation loop with automated quality control, and it opens avenues for scaling and transferability of discovered patterns.

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.

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

TL;DR

We address generalization gaps in small language models by introducing PaDA-Agent, an evaluation-driven data augmentation framework that links validation-time failures to targeted synthetic data generation. A central orchestrator coordinates Pattern Analysis, Data Generation, and Quality Control to discover generalization patterns, draft augmentation strategies, and ensure data utility. Empirical results on a 1B Instruct Llama model show consistent improvements across reasoning, knowledge, and coding tasks, with pronounced gains in low-data regimes and interpretable strategies that reveal underlying failure modes. The approach reduces manual data engineering by closing the generalization gap through a principled, evaluation-informed augmentation loop with automated quality control, and it opens avenues for scaling and transferability of discovered patterns.

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.
Paper Structure (23 sections, 9 equations, 8 figures, 3 tables, 1 algorithm)

This paper contains 23 sections, 9 equations, 8 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. PaDA-Agent Architecture
  4. Pattern Analysis Agent
  5. Data Generation Agent
  6. Quality Control and Efficiency
  7. Experimental Results
  8. Datasets
  9. Experiment Setup
  10. Results
  11. Ablation Studies
  12. Conclusion and Future Work
  13. Detailed metholodgy
  14. PaDA-Agent Architecture
  15. Pattern Analysis Agent
  16. ...and 8 more sections

Figures (8)

  • Figure 1: The architecture of PaDA-Agent: The Central Orchestrator coordinates three specialized agents: the Pattern Analysis Agent (for error analysis, pattern categorization, and augmentation drafting), the Data Generation Agent (for pattern-guided and diverse synthetic data creation), and the Quality Control Agent (for adherence, utility, and relevancy checks). This coordinated process enables improved fine-tuning of SLMs with high-quality, targeted data augmentation.
  • Figure 2: Prompt for Sample-Level Error Analysis Subagent.
  • Figure 3: Prompt for Pattern Categorization Subagent.
  • Figure 4: Prompt for Strategy Drafting Subagent.
  • Figure 5: Prompt for Data Generation Agent - Learn from Generalization Patterns.
  • ...and 3 more figures