Effect of Synthetic Data on Indic LLM Performance

IndicLLMSuite extensively leverages synthetic data to address resource scarcity for Indian languages, including large-scale machine translation of English Wikimedia into 14 Indian languages, transliteration to Roman script, and LLM-generated instruction-following conversations. While this increases data coverage and diversity, the authors acknowledge that translated and synthetic content may not fully capture real-world language usage, potentially affecting model naturalness and generalization.

The authors explicitly defer a systematic study of how synthetic data influences model performance, leaving open questions about optimal proportions, domains, and types of synthetic data, as well as their interaction with verified and unverified sources in pre-training and fine-tuning mixtures.