Datasets:

surrey-nlp
/

dialect-preferences

license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - dialect
  - english-varieties
  - dpo
  - preference-dataset
size_categories:
  - 10K<n<100K

DiaLLM — Pooled Preference Dataset (Implicit Thread)

Part of DiaLLM: An Investigation into the Robustness-Generation Gap in English Dialect Adaptation (EMNLP 2026 Main).

45,690 preference pairs, pooling all three variety-specific sets (Australian, Northern British, Indian) without variety targeting. Used for implicit-thread DPO training, where the three varieties are pooled rather than targeted individually, preserving the variety-agnostic objective of that thread.

Construction

Built from the UltraFeedback preference dataset (Cui et al., 2023), via Argilla's cleaned/binarized release (argilla/ultrafeedback-binarized-preferences-cleaned, MIT licensed): the originally-preferred completion is transformed into a dialectal variant using Multi-VALUE (Ziems et al., 2023), based on eWAVE morphosyntactic features. Code blocks are preserved verbatim during conversion.

Columns

Column	Description
`prompt`	Original UltraFeedback prompt (standard English, unmodified)
`chosen`	Multi-VALUE dialectal variant of the preferred completion
`rejected`	Original standard-English form of that completion
`prompt_dialect_density`	eWAVE-based dialect feature density of the prompt
`chosen_dialect_density`	eWAVE-based dialect feature density of the chosen completion

Code, checkpoints, linguistic-analysis toolkit: https://github.com/surrey-nlp/diallm

Paper: https://arxiv.org/abs/2607.07669

Citation

@article{painter2026diallm,
  title     = {DiaLLM: An Investigation into the Robustness-Generation Gap in English Dialect Adaptation},
  author    = {Painter, Jordan and Srirag, Dipankar and Kappiyath, Adarsh and Kanojia, Diptesh and Joshi, Aditya and Yin, Lu},
  year      = {2026},
  eprint    = {2607.07669},
  archivePrefix = {arXiv}
}