SynthLLM: An LLM-based Scalable Synthetic Data Generation Pipeline for Low-Resource Languages

Abstract

Large Language Models (LLMs) have enabled scalable synthetic data generation, yet their effective adaptation to low-resource languages remains underexplored. We introduce an LLM-based generate and annotate paradigm to create synthetic datasets for low-resource NLP classification tasks. The framework employs a smaller model for text generation and a stronger model for automatic annotation. Using Farsi Natural Language Inference (NLI) as a case study, we construct a large-scale synthetic dataset of 100,000 labeled instances. We provide a systematic empirical analysis of annotation quality, label-distribution effects, and training regimes. We compare GPT-4o-mini, Aya-23-35B, and DeBERTa as annotators and examine how annotation variability propagates to downstream performance. Our results show that a warm-up phase with synthetic data consistently outperforms data mixing and reversed ordering. Notably, open-source annotation (Aya-23-35B) achieves comparable downstream performance to the proprietary model (GPT-4o-mini), with significant cost implications for deploying pipelines in low-resource settings. The dataset and code are publicly available at https://huggingface.co/datasets/Solmazp/text2entail.

Resources

Details