A fine-tuned
DeBERTa-v3-base model for classifying English text into three dialect varieties:
en-AU (Australian),
en-IN (Indian), and
en-UK (British).
Trained as part of the DiaLLM project — a study of dialect-adapted language models using CPT, SFT, DPO, GRPO, and GSPO across Gemma, Llama, and Qwen model families. Used as an independent evaluation metric to assess whether generated text exhibits target-dialect characteristics.
1from transformers import pipeline
2
3classifier = pipeline(
4 "text-classification",
5 model="jordanpainter/diallm-dialect-classifier",
6)
7
8classifier("I reckon it's a ripper idea, mate.")
9# [{'label': 'en-AU', 'score': 0.87}]
Fine-tuned on
BESSTIE-CW-26, a dataset of 6,243 naturally occurring English sentences annotated for dialect variety. All splits were pooled and re-split 80/10/10 with stratification to ensure balanced dialect representation in dev and test.
Indian English is the most separable class; Australian and British English share substantial lexical overlap, leading to some inter-class confusion between the two.
