From Seed to Synthetic: A Data Generation Protocol for Fair, Inclusive, and Robust Language Detection

High-quality annotated data remains scarce for NLP in low-resource, fairness-sensitive settings. Large language models (LLMs) enable scalable synthetic data generation, but systematic control of label quality, diversity, and fairness is lacking. We propose From Seed to Synthetic, a seven-stage protocol combining controlled generation workflows, multi-dimensional quality metrics (diversity, completeness, bias amplification), and fairness-aware evaluation (Equalized Odds, subgroup FPR/FNR) into a reusable blueprint.We validate it on two Italian case studies sharing the same evaluation stages but different generation workflows: inclusive-language detection in job ads (masking-and-replacement) and gender-bias detection in machine-translated sentences (translation-and-paraphrasing). Fine-tuning Phi-3.5-mini-instruct on 8,500 augmented instances raises macro F1 by up to 30 points over pre-trained baselines in the first study. In the harder second study, built on an extremely imbalanced seed (1% inclusive labels), pre-trained models perform near chance while fine-tuning achieves the best balanced accuracy and cuts a bias-amplification metric from 0.75 to about 0.56, though detecting the minority inclusive class remains difficult. Both studies show fine-tuning shifting gender-related bias distributions, underscoring the need for continuous fairness monitoring. The protocol offers a replicable, fairness-aware foundation for synthetic data generation in gendered, low-resource languages.