Transfer of LoRA Gains to Dialect-to-Standard Normalization

Determine whether parameter-efficient LoRA fine-tuning transfers the gains observed for dialect-to-English translation to dialect-to-Standard-Bangla normalization when using the 5-Dialects-BN benchmark.

Background

The paper applies LoRA fine-tuning to dialect-to-English translation and subjectivity classification, using 160 training instances per dialect, but does not fine-tune LoRA adapters for dialect-to-Standard-Bangla normalization. Normalization is evaluated primarily through prompting, with the only supervised comparison provided by fine-tuned IndicBART and BanglaT5 models. Because the reported supervised experiments do not test LoRA on normalization, it remains unresolved whether the strong translation-side improvements produced by LoRA would also improve dialect-to-Standard-Bangla normalization.

References

Our adaptation experiments cover translation and subjectivity classification; dialect-to-standard normalization is benchmarked under prompting only, and the sole supervised normalization comparison we report is the Indic seq2seq fine-tune in Appendix~\ref{app:indic_baselines}, which suggests that $160$ dialect-aligned examples are not sufficient for a small Indic seq2seq model to match strong prompting on this task; the extent to which LoRA transfers the translation-side gains to normalization remains open.

— 5-Dialects-BN: Unmasking the Impact of Transliteration on Bangla Dialectal LLMs  (2609.09964 - Jawad et al., 9 Sep 2026) in Scope of the reported experiments, Section "Limitations"