AI Research Portal

This paper introduces DiaLLM, a framework for adapting large language models (LLMs) to English dialects. It addresses the gap in dialectal generation by employing a full-pipeline approach that includes continual pretraining, supervised fine-tuning, and various alignment strategies. The study finds that dialectal robustness and generation are distinct, with benchmarks primarily reflecting pretraining and fine-tuning, while alignment influences generation in ways not captured by benchmarks. Explicit, variety-targeted adaptation leads to outputs recognized and preferred as dialectal, but the method optimizing the dialectal reward does not necessarily yield the most preferred outputs.

The paper addresses the following problems:

Large language models (LLMs) often fail to generate text in dialectal English, despite understanding it. This leaves the more challenging task of dialectal generation largely unaddressed.
Existing research on dialectal adaptation primarily focuses on robustness to dialectal input (Natural Language Understanding) rather than generation of dialectal output.
Current methods for dialectal adaptation are often limited to specific stages of the LLM pipeline (e.g., adapters for NLU) and do not investigate a full-pipeline approach encompassing pretraining, fine-tuning, and alignment for generation.
There is a lack of systematic comparison of different adaptation strategies and alignment methods across various English dialects, particularly regional ones like Australian, Indian, and Northern British English.

DiaLLM: A full-pipeline dialect adaptation framework that includes continual pretraining, supervised fine-tuning, and three alignment methods (DPO, GRPO, GSPO), with a controlled comparison across two post-training paradigms and three open-weight LLM families.
Establishment of a robustness-generation gap: Benchmark performance is primarily determined by continual pretraining (CPT) and supervised fine-tuning (SFT), while alignment methods subtly reshape generation in ways not captured by benchmarks.
Exposure of a reward-quality gap: Explicit, variety-targeted adaptation is recognized as dialectal and preferred, but the method that most aggressively optimizes the dialectal reward (GRPO) is not preferred by human evaluators or LLM judges, indicating a mismatch between reward optimization and perceived dialectal quality.

The DiaLLM framework involves a multi-stage process for dialect adaptation:

Continual Pretraining (CPT): Three open-weight LLM families (Llama 3.1-8B, Qwen 3-8B, Gemma 3-4B-it) are continually pretrained on the International Corpus of English (ICE), which contains 18 English varieties. This stage introduces dialectal exposure.
Post-Training Paradigms: Two paradigms are pursued:
- Implicit Thread: This involves standard instruction-tuning and broad alignment, without specific variety targeting.
- Explicit Thread: This uses dialectally-perturbed supervision and variety-targeted alignment, focusing on specific English varieties (Australian, Indian, Northern British).
Supervised Fine-Tuning (SFT): Both threads apply SFT from the CPT checkpoint. The implicit thread uses a standard ultrafeedback corpus for general instruction following. The explicit thread uses a dialect-specific dataset, preserving the dialectal signal introduced by CPT.
Alignment Strategies: Three preference optimization methods are compared:
- Direct Preference Optimisation (DPO): An offline method that optimizes directly from contrastive pairs.
- Group Relative Policy Optimisation (GRPO) and Group Sequence Policy Optimisation (GSPO): These methods optimize a shared reward function that combines dialectal feature density and meaning preservation.
Dialect Feature Classifier: A multi-label classifier is trained to detect 135 morphosyntactic and lexical features of dialectal English, derived from the eWAVE typology. This classifier provides a reward signal for online alignment.
Reward Formulation: The reward function for GRPO and GSPO is defined as R = λ·dial(y) + (1-λ)·comet(y, ŷ) + (1-λ)·cos(y, ŷ), where dial(y) is the dialectal reward from the feature classifier, and comet and cos measure semantic fidelity.

Base Models: Llama 3.1-8B, Qwen 3-8B, and Gemma 3-4B-it.
Continual Pretraining Corpus: International Corpus of English (ICE), comprising approximately 20 million tokens from 18 varieties.
Post-Training Dataset: Preference dataset from Cui et al. (2023), converted into dialectal variants using the Multi-VALUE library for Australian English, Indian English, and Northern British English. Dataset sizes: 11.8k (en-AU), 15.4k (en-IN), 18.4k (en-UK) preference pairs.
Dialect Feature Classifier: BERT-base encoder trained on ~148,000 sentences, detecting 135 morphosyntactic and lexical features.
Alignment Methods: DPO, GRPO, and GSPO.
Reward Function: R = λ·dial(y) + (1-λ)·comet(y, ŷ) + (1-λ)·cos(y, ŷ), with λ fixed at 0.80.
Training Settings:
- CPT: GaLore, AdamW, 64,000 steps, learning rate 1 × 10⁻⁴, effective batch size 512.
- SFT: 3 epochs, learning rate 2 × 10⁻⁵, cosine scheduling.
- DPO: 5,000 steps, sigmoid loss, β = 0.1.
- GRPO/GSPO: 5,000 steps, sampling temperature 0.7, top-p = 0.95, β = 0.02, € = 3 × 10⁻⁴, high = 4 × 10⁻⁴. Reward weights (0.8, 0.1, 0.1).
Infrastructure: NVIDIA A100 GPUs, SLURM-managed cluster, accelerate library, bfloat16 precision, gradient checkpointing, fixed random seed 1234.

Datasets: VALUE, BESSTIE (BST-Sent, BST-Sarc), DIALECTBENCH (DID), GLUE, GPQA, BBH.
Metrics: Variety classifier accuracy, human preference evaluation, LLM-as-judge evaluation, dialectal feature density, diversity, and stacking.
Baselines: Base models, CPT, SFT.
Numeric Results:
- In the explicit thread, SFTd generally sets a strong baseline, with alignment methods producing small, inconsistent shifts.
- Explicit, variety-targeted adaptation is preferred by human annotators and LLM judges over broad alignment for en-IN and en-UK.
- GRPO, which maximizes the eWAVE reward density, produces the fewest independent surface dialectal markers for Llama and Qwen.
- Human preference tracks dialectal marking but not reward density; more dialectal output is preferred over the standard baseline, but more natural output is preferred over GRPO.
- Indian English (en-IN) shows the highest surface dialectal marking.
Runtime/Training Cost: Not stated in the paper.

The VALUE dataset was developed for AAVE and is not specifically designed for the three target varieties, meaning gains are interpreted as broad dialectal robustness rather than variety-specific improvement.
BESSTIE covers the target varieties but is limited to sentiment and sarcasm, not assessing broader dialectal competence.
eWAVE was designed as a typological database, not a reward signal for generative training, leading to potential mismatches where features are pragmatically conditioned or vary at the individual speaker level.
The reward function and generation evaluation share the eWAVE inventory, meaning feature density gains reflect alignment with the training space rather than independently verified authentic dialect use.
The BESSTIE-trained variety classifier exhibits biases and near-total collapse to en-AU predictions for en-UK LLM-generated outputs, indicating it does not generalize well to generated text.
The ICE corpus primarily consists of formal and semi-formal texts, potentially limiting generalization to informal or social media registers.
The pipeline is limited to English dialects.
Generation evaluation was conducted only on Llama-3.1-8B, and results may not generalize to other models.
Human preference evaluation used a small pool of annotators, and inter-annotator agreement was modest, particularly for en-AU.
The implicit and explicit adaptation threads differ in SFT training data (UltraFeedback vs. dialect-specific preference dataset), making it difficult to isolate the effect of variety targeting alone.

Method evidence: "We introduce DiaLLM, a dialect adaptation framework that continually pretrains three open-weight LLM families (Llama 3.1-8B, Qwen 3-8B, Gemma 3-4B) on the International Corpus of English (ICE)... from each CPT checkpoint we pursue two post-training paradigms: implicit adaptation applies standard instruction-tuning and broad alignment; explicit adaptation uses dialectally-perturbed supervision and variety-targeted alignment." (Section: 1 Introduction)
Result evidence: "Our results reveal a dissociation between dialectal robustness and dialectal generation: benchmark performance is driven primarily by CPT and SFT, while alignment visibly shapes generation in ways benchmarks do not capture." (Section: 1 Introduction)
Limitation evidence: "Our evaluation has several limitations. First, VALUE was developed around AAVE and is not designed to benchmark our three target varieties directly; gains are best interpreted as evidence of broad dialectal robustness rather than variety-specific improvement." (Section: Limitations and Ethical Considerations)

DiaLLM: An Investigation into the Robustness-Generation Gap in English Dialect Adaptation

Abstract

Summary

Problem

Key Contributions

Method

Technical Details

Results

Weaknesses / Limitations

Additional Figures