Part of DiaLLM: An Investigation into the Robustness-Generation Gap in
English Dialect Adaptation (EMNLP 2026 Main).
18,402 preference pairs for Indian English (en-IN), used for explicit-thread
DPO/GRPO/GSPO training targeting this variety.
Built from the UltraFeedback preference dataset (Cui et al., 2023):
the originally-preferred completion is transformed into a dialectal variant
using Multi-VALUE (Ziems… See the full description on the dataset page:
https://huggingface.co/datasets/surrey-nlp/alignment-indian-final.
