DiaLLM: An Investigation into the Robustness-Generation Gap in English Dialect Adaptation Jordan Painter1 Dipankar Srirag2 Adarsh Kappiyath1 Diptesh Kanojia1 Aditya Joshi2 Lu Yin1 1
Institute for People-Centered AI, University of Surrey, Surrey, United Kingdom 2 University of New South Wales, Sydney, Australia {j.painter,a.kappiyath,d.kanojia,l.yin}@surrey.ac.uk {d.srirag,aditya.joshi}@unsw.edu.au
arXiv:2607.07669v1 [cs.CL] 8 Jul 2026
Abstract Large language models increasingly understand dialectal English, yet still produce only standard, US-leaning English, leaving dialectal generation, the harder half of the problem, largely unaddressed. We introduce DiaLLM, which continually pretrains three open-weight language model families on the International Corpus of English and applies implicit and explicit post-training paradigms, each combined with three model alignment strategies, giving the first controlled comparison of these components across Australian, Indian, and Northern British English. Our results reveal that dialectal robustness and generation are dissociated: benchmarks are shaped by continual pretraining and SFT, while alignment visibly reshapes generation in ways benchmarks do not capture. Explicit variety-targeted adaptation produces output reliably recognised as dialectal and preferred over broad alignment, yet the method that most aggressively optimises the dialectal reward is not preferred by human evaluators. Independent linguistic analysis corroborates this reward-quality gap, most clearly on two of the three families. No single alignment method dominates, and closing the gap will require richer reward designs and continued investment in dialectal resources. We release all code, checkpoints, and preference datasets.
1
Figure 1: Overview of the DiaLLM pipeline: continual pretraining on ICE followed by either implicit adaptation (standard SFT + alignment) or explicit adaptation (dialectal SFT + variety-targeted alignment), with DPO, GRPO, and GSPO compared across both paradigms.
(Gururangan et al., 2020), robustness to dialectal variation within a single language remains comparatively underexplored. Existing work is concentrated heavily on AAVE and code-switching (Blodgett et al., 2016; Sap et al., 2019; Hofmann et al., 2024), while regional varieties such as Australian, Indian, and Northern British English receive substantially less attention at the alignment stage. Methods that do address dialectal variation— TADA (Held et al., 2023), HyperLoRA (Xiao et al., 2023)—do so via dialect adapters or low-rank components targeting NLU robustness under dialectal input; LoRDD (Srirag et al., 2025a) similarly employs a low-rank dialect adapter for decoder models in a task-specific setting. None of them investigates full-pipeline adaptation across pretraining, fine-tuning, and alignment, nor whether models can produce dialectally appropriate output. Dialectal generation (producing a language variety on output, not merely tolerating it on input) is
Introduction
Large language models (LLMs) achieve strong performance across a range of NLP tasks (Wang et al., 2024; Touvron et al., 2023), but often falter on dialectal and non-standard varieties—regional and social forms that diverge from standardised norms (Joshi et al., 2025; Blodgett et al., 2016; Hofmann et al., 2024; Mire et al., 2025). These disparities reflect structural imbalances in training data: large-scale corpora privilege majority language forms, marginalising non-standard varieties (Bender et al., 2021; Joshi et al., 2020). While domain adaptation has received substantial attention 1
1. DiaLLM, a full-pipeline dialect adaptation framework with a controlled comparison of continual pretraining, supervised fine-tuning, and three alignment methods (DPO, GRPO, GSPO) across two post-training paradigms and three open-weight LLM families. We release all code (including a linguistic-analysis toolkit), model checkpoints, and dialectal preference datasets1 .
a critical part of the problem that current methods do not address. Consider three users chasing a late parcel: Indian English: My parcel is not yet coming; kindly do the needful and prepone the delivery. Northern British English: Me parcel’s not arrived; summat were up wi’ the courier.
2. We establish a robustness–generation gap: benchmark performance is set by CPT and SFT, whereas alignment moves benchmarks only marginally, yet visibly reshapes generation, showing benchmarks alone misjudge what alignment does.
Australian English: Me parcel still hasn’t turned up, reckon it’s good this arvo, eh? A modern LLM will understand all three, yet answer each in flattened Standard American English. A model that understands every variety but speaks only one has not solved the dialect problem; it has hidden half of it. We introduce DiaLLM, a dialect adaptation framework that continually pretrains three openweight LLM families (Llama 3.1-8B, Qwen 3-8B, Gemma 3-4B) on the International Corpus of English (ICE; Greenbaum and Nelson 1996), spanning 18 English varieties. As shown in Figure 1, from each CPT checkpoint we pursue two posttraining paradigms: implicit adaptation applies standard instruction-tuning and broad alignment; explicit adaptation uses dialectally-perturbed supervision and variety-targeted alignment. Across both paradigms, we compare Direct Preference Optimisation (DPO), Group Relative Policy Optimisation (GRPO), and Group Sequence Policy Optimisation (GSPO). We treat the relative effectiveness of each pipeline component and alignment method as an empirical question rather than assuming any single recipe is best. 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. Explicit varietytargeted adaptation is reliably recognised as dialectal and preferred over broad alignment by both human annotators and LLM judges. Yet, the method that most aggressively optimises the dialectal reward is not the one they prefer. Our linguistic analysis confirms that this method produces the fewest recognisable dialectal markers across all three families, showing a mismatch between feature-density reward and perceived dialectal quality. We make three primary contributions:
3. We expose a reward–quality gap: although explicit, variety-targeted adaptation is reliably recognised as dialectal and preferred over broad alignment, the method that most aggressively optimises the dialectal reward (GRPO) is not the one LLM judges or human annotators prefer on Llama-3.1-8B, where preference evaluation is conducted; an independent linguistic analysis across all three families corroborates this finding.
2
Related Work
LLMs systematically underperform on dialectal and non-standard English: standard NLP tools fail on AAVE (Blodgett et al., 2016), dialectal text is disproportionately flagged as toxic (Sap et al., 2019), and LLMs encode racio-linguistic stereotypes and generate lower-quality outputs for dialectal prompts even after alignment (Hofmann et al., 2024; Mire et al., 2025), with reward models themselves steering outputs toward White Mainstream English (Mire et al., 2025). These gaps are documented at scale across 281 varieties using VALUE (Ziems et al., 2022), BESSTIE (Srirag et al., 2025b), and DialectBench (Faisal et al., 2024). Methods that do address dialectal variation— TADA (Held et al., 2023) and HyperLoRA (Xiao et al., 2023)—do so via dialect adapters or lowrank components targeting NLU robustness under dialectal input; LoRDD (Srirag et al., 2025a) similarly employs a low-rank dialect adapter for decoder models in a task-specific setting. None of 1
2
https://github.com/jordanpainter/diallm
these existing works investigates full-pipeline adaptation across pretraining, fine-tuning, and alignment paradigms, nor whether models can produce dialectally appropriate output. Continual pretraining on in-domain text is a well-established adaptation foundation (Gururangan et al., 2020), applied across medical, legal, and multilingual settings (Xie et al., 2024; Huang et al., 2023); Khan et al. (2025) show dialect gaps in Québec French can be narrowed via CPT with limited benchmark degradation. For posttraining, DPO (Rafailov et al., 2024), GRPO (Shao et al., 2024), and GSPO (Zheng et al., 2025) offer contrasting approaches to preference optimisation; De Langis et al. (2024) demonstrate that RL with automatic linguistic reward signals can steer generation toward stylistically defined targets—a principle we extend to dialect-specific features. None of these methods have been systematically compared for dialect adaptation.
3
Methodology
3.1
Base Models
varieties was sought but could not be obtained, either because the relevant ICE teams did not respond to requests or because the data was not in a usable digitised form. The corpus covers a broad range of regional, cultural, and postcolonial Englishes, including Hong Kong, Nigeria, India, Jamaica, and others. In total, our training corpus comprises approximately 20 million tokens. Pre-training setup: We continually pretrain each selected model on the cleaned ICE corpus before any supervised or preference-based posttraining using GaLore (Zhao et al., 2024), enabling memory-efficient full-parameter optimisation. CPT is implemented as causal language modelling over the cleaned ICE text; full training details are provided in Appendix C. 3.3
For post-training, we use the preference dataset introduced by Cui et al. (2023), which contains highquality prompt-response pairs annotated with binary preferences over completions. For the dialectspecific training thread, we convert the preferred completions into dialectal variants using the MultiVALUE library (Ziems et al., 2023), targeting three English varieties: Australian English, Indian English, and Northern British English. Multi-VALUE applies transformations based on eWAVE features (Kortmann et al., 2020), which constitute the only feature-level inventory of morphosyntactic variation spanning all English varieties considered in this work. This makes eWAVE a natural foundation for both dialect perturbation and reward specification. During conversion, code blocks are automatically detected and preserved verbatim. The resulting datasets contain approximately 11.8k, 15.4k, and 18.4k preference pairs for Australian, Northern British, and Indian English, respectively. Differences in dataset size arise from unsuccessful conversions produced by Multi-VALUE.
We apply our dialect adaptation pipeline to three open-weight LLM families spanning the 4B–8B parameter range: Llama 3.1-8B (Grattafiori et al., 2024), Qwen 3-8B (Yang et al., 2025), and Gemma 3-4B-it (Team et al., 2025)2 . Selecting architecturally diverse models allows us to assess whether gains from dialect adaptation are robust across training regimes and vocabulary choices, rather than specific to a single architecture. 3.2
Post-Training Dataset
Continual Pretraining
Corpus: We continually pretrain all models using the International Corpus of English (ICE), a collection of comparable corpora from over 25 English-speaking regions. Each variety contains one million words of spoken and written language sampled using consistent sociolinguistic criteria, making ICE uniquely suited for this task: it is the only available multi-variety corpus spanning both inner and outer circle varieties at a usable scale, with no comparable resource covering this range of regional Englishes. For this work, we include 18 ICE varieties with cleaned, digitised data available.3 Access to further
3.4
Alignment Strategy
From each CPT checkpoint, we pursue two dialect adaptation paradigms: the implicit thread (broad training, with no variety targeting) and the explicit thread (single-variety targeting throughout), as illustrated in Figure 1. We treat each stage as serving a distinct role in the pipeline: continual pretraining introduces dialectal exposure that
2
For Gemma 3-4B-IT, we initialise CPT from the instruction-tuned variant rather than the base model, following poor performance in preliminary experiments with the base checkpoint; the remaining pipeline is identical. 3 See Appendix A for preprocessing and filtering details.
We could only obtain permissions to use the corpus for this set.
3
P ϕdial (y) = log(1 + i∈F σ(logiti )), where F is the active feature set: all 135 features in the implicit thread, or the attested variety-specific subset in the explicit thread (en-AU: 20 features; en-UK: 30 features; en-IN: 38 features).
is otherwise absent from models trained predominantly on standard English corpora; post-training methods can shape output distributions but cannot introduce dialectal signal that is not already latent in the model’s weights; supervised fine-tuning restores general instruction-following behaviour.
3.4.4 Reward Formulation Both GRPO and GSPO optimise a shared reward combining dialectal feature density and meaning preservation:
3.4.1 Supervised Fine-Tuning Both adaptation threads apply SFT from the CPT checkpoint, training only on chosen completions in a single-turn instruction format with early stopping based on held-out evaluation loss. Full hyperparameters are reported in Appendix C. The implicit thread fine-tunes on the standard ultrafeedback-binarized-preferences corpus, providing a general instruction-following objective with no explicit dialectal signal. The explicit thread fine-tunes instead on the dialect-specific dataset for the target variety, so the chosen completions already contain Multi-VALUE dialectal features. No general SFT stage intervenes between CPT and dialect SFT in the explicit thread, preserving the dialectal signal introduced by CPT.
1−λ R = λ·ϕdial (y)+ 1−λ 2 ·ϕcomet (y, ŷ)+ 2 ·ϕcos (y, ŷ) (1) ϕdial (y) is the log-sum dialectal reward from the feature classifier (Section 3.4.3). ϕcomet (y, ŷ) and ϕcos (y, ŷ) measure semantic fidelity against the SFT chosen completion ŷ using COMET (Rei et al., 2022) and cosine similarity over all-MiniLM-L6-v2 (Reimers and Gurevych, 2019) respectively. We fixed λ = 0.80 across all model families and alignment methods; full selection details are given in Appendix C. All components are z-score normalised.
3.4.5 GRPO and GSPO GRPO (Shao et al., 2024) generates a group of rollouts per prompt, estimates a baseline from the group mean reward, and computes token-level policy gradient updates without a value network. GSPO (Zheng et al., 2025) replaces token-level importance sampling with sequence-level reweighting, assigning credit by the full-sequence probability ratio under the current and reference policies, avoiding the bias introduced by per-token clipping. Both use four completions per prompt and the reward in Section 3.4.4.
3.4.2 Direct Preference Optimisation DPO (Rafailov et al., 2024) is an offline method that optimises directly from contrastive pairs without querying a reward model. Preference pairs are drawn from the dialectal datasets in Section 3.3: chosen completions are Multi-VALUE dialectal variants; rejected completions are the original standard English forms. In the implicit thread, DPO trains on the combined three-variety dataset; the three varieties are pooled rather than targeted individually, preserving the variety-agnostic objective of this thread. In the explicit thread, only the targetvariety subset is used. 3.4.3 Dialect Feature Classifier To provide a reward signal for online alignment, we train a multi-label classifier to detect 135 morphosyntactic and lexical features of dialectal English derived from the eWAVE typological database (Kortmann and Lunkenheimer, 2013). eWAVE provides the only structured feature-level typology of morphosyntactic variation spanning our three target varieties, making it the sole available basis for a linguistically interpretable reward signal at this level of dialectal specificity. Full architecture, training, and calibration details are given in Appendix D. The classifier’s logits are passed through a sigmoid and the resulting probabilities are aggregated as
4
Experiment Setup
4.1
General Benchmarks
We evaluate general capability using BBH (Suzgun et al., 2022), GPQA (Rein et al., 2023), and GLUE (Wang et al., 2018), implemented via the EleutherAI/lm-evaluation-harness (Gao et al., 2023) with unmodified task templates. 4.2
Dialectal Benchmarks
We evaluate dialectal robustness using three benchmark suites. The dialect identification (DID) task from D IALECT B ENCH (Faisal et al., 2024) requires classifying inputs as British English (ENGB), American English (EN-US), or generic English (EN), with higher scores indicating greater 4
Qwen3-8B Benchmark
SFTd
Llama 3.1-8B
Gemma 3-4B-it
DPO
GRPO
GSPO
SFTd
DPO
GRPO
GSPO
SFTd
DPO
GRPO
GSPO
en-AU (Australian English) GLUE 61.59 61.08 VALUE 55.94 50.00 BST-Sent 90.02 89.57 BST-Sarc 37.07 37.36 DID 67.11 65.44
61.60 55.04 89.93 36.84 66.61
61.65 55.59 89.92 36.52 64.94
46.26 45.42 87.31 39.45 67.95
47.11 47.06 87.26 41.18 64.44
47.08 45.75 87.35 40.54 67.11
47.31 45.61 87.21 38.95 69.28
52.69 50.59 84.91 27.31 35.56
53.39 47.41 84.73 31.86 35.73
51.98 49.28 84.91 28.90 35.56
52.95 50.60 84.76 28.30 35.56
en-IN (Indian English) GLUE 62.71 VALUE 52.92 BST-Sent 89.75 BST-Sarc 39.27 DID 71.45
62.50 52.94 89.78 39.43 70.79
62.84 52.90 90.01 38.87 70.45
54.27 48.32 81.52 36.36 64.78
56.45 49.34 80.95 37.68 65.78
53.94 47.88 82.64 35.47 64.11
53.88 48.01 82.09 35.27 63.44
57.15 50.21 77.50 32.79 13.02
57.63 49.49 75.53 32.54 13.02
57.12 50.68 78.97 32.88 13.02
58.38 50.78 79.11 33.22 13.36
en-UK (Northern British English) GLUE 63.48 63.42 63.99 VALUE 58.27 54.59 58.50 BST-Sent 87.92 88.01 88.01 BST-Sarc 35.44 34.97 35.28 DID 63.94 63.44 64.27
63.72 58.49 87.99 35.48 64.61
54.05 45.67 79.45 34.96 50.58
53.77 48.35 80.83 34.58 58.60
53.95 45.56 80.11 35.16 49.58
53.78 45.59 79.89 34.73 51.59
56.03 48.93 82.53 33.95 35.23
55.35 46.36 79.31 31.68 35.39
55.45 48.37 82.23 33.11 35.23
56.42 48.61 83.49 32.61 35.23
63.03 50.41 89.61 38.50 70.12
Table 1: Explicit thread alignment comparison (%). SFTd denotes dialect-specific supervised fine-tuning; DPO, GRPO, and GSPO are applied with variety-targeted rewards. Bold indicates the best result within each model family per row and dialect group. BST-Sent and BST-Sarc refer to BESSTIE sentiment and sarcasm respectively.
sensitivity to dialectal variation; as DID contains no en-AU or en-IN class, scores for these conditions index general dialectal sensitivity rather than variety-specific recognition. VALUE (Ziems et al., 2022) replaces GLUE inputs with AAVE variants, directly measuring performance disparities relative to the standard benchmark. BST-Sarc and BSTSent from BESSTIE (Srirag et al., 2025b) evaluate sarcasm detection and sentiment classification across British, Indian, and Australian English, and constitute the only available coverage specific to our three target varieties among existing English dialect benchmarks. 4.3
and personal reflection, designed to elicit natural, conversational responses (Appendix H.2). All evaluation methods use this same prompt set. Automatic Metrics. Variety classifier accuracy (Appendix F) is the proportion of responses correctly classified as the target variety by a DeBERTav3-base model fine-tuned on a deduplicated, curated split of BESSTIE (Srirag et al., 2025b) using an 80/10/10 stratified train/validation/test split, achieving a macro F1 of 0.766 on the held-out test set. It serves as our primary automatic proxy for holistic dialectal style. Human Preference Evaluation. Two native or near-native speakers per dialect (six annotators total), selected on the basis of self-reported familiarity with the target variety, complete 100 blind pairwise or three-way judgements structured across four tasks of 25 trials each, comparing successive pipeline stages and alignment methods (detailed in Section 5.4). Trials are presented in blocks; model identities are not revealed. Tasks 1, 2, and 3 use A/B/Tie judgements; Task 4 is a forced three-way choice with no tie option.
Generation Evaluation
Benchmarks measure performance under dialectal input but do not reveal whether models produce dialectally marked output. We, therefore, conduct a dedicated generation evaluation using Llama-3.1-8B, which shows the most consistent dialect-sensitivity gains across pipeline stages and allows detailed qualitative examination than averaging across families would permit. We focus on the explicit thread for variety-targeted generation analysis, with the broad thread also included where alignment methods are compared directly (Section 5.4).
LLM-as-Judge Evaluation. We evaluate the same four tasks using microsoft/phi-4 (Abdin et al., 2024), a model outside the evaluated families to reduce architectural bias. Both conditions instruct the judge to focus on dialectal features and ignore fluency and writing quality. They differ only
Prompts. We construct a set of 25 open-ended casual prompts spanning five domains: casual conversation, opinion, food and lifestyle, hypothetical,
5
in the system prompt: no-features, where Phi-4 acts as a dialect specialist without an explicit feature list; and eWAVE-features, where the system prompt additionally provides the full eWAVE feature inventory for the target variety. Comparing conditions isolates the contribution of explicit linguistic knowledge to LLM dialectal judgements.
5
Results
5.1
Implicit and Explicit Adaptation
roborates this pattern across all three families: the method that maximises the eWAVE reward (GRPO) produces the fewest surface dialectal markers and the lowest feature diversity for Llama and Qwen (Appendix J, Table 11; Figure 2). This is in part a property of the reward basis itself: eWAVE is a typological atlas documenting features that are attested across varieties, not a perceptual evaluation instrument. Some dialectal features overlap across multiple varieties, and others carry stronger associations with non-target varieties in general language use, meaning reward optimisation over this feature set can increase surface feature density without producing outputs that are holistically recognisable as the target variety.The en-UK classifier does not generalise to generated outputs (Appendix F), so en-UK generation quality is assessed via human and LLM judgements only.
In the implicit thread (full results in Appendix G), CPT disrupts general benchmark performance, SFT accounts for the majority of subsequent recovery, and alignment produces inconsistent shifts with no method reliably improving over SFT. Table 1 shows that in the explicit thread, SFTd generally sets a strong performance baseline and alignment methods produce mostly small, inconsistent shifts. General capability benchmarks remain stable across all methods, suggesting alignment does not degrade general capability in the explicit thread. DPO shows the widest variance across varieties, with a consistent tendency to hurt on VALUE. Across varieties, no single method dominates and benchmark differences alone do not cleanly distinguish the alignment approaches, motivating the subsequent generation quality analysis in Section 5.2 onwards. 5.2
5.4
Table 3 reports preference results from human annotators (two per locale) alongside two Phi-4 judge conditions, evaluated on Llama-3.1-8B outputs throughout. The clearest finding across evaluators is that explicit adaptation produces output that human annotators recognise as dialectally distinct. Annotators prefer SFTd over the instruction-tuned baseline at high rates for en-IN and en-UK (84% and 87% respectively, ties excluded), and prefer explicitly variety-targeted alignment over broad alignment at similarly high rates (71% en-IN, 85% en-UK). Both Phi-4 conditions agree unanimously on the latter comparison across all three varieties. These results indicate that the explicit pipeline produces perceptible dialectal features, and that variety targeting produces more recognisable output than broad alignment regardless of evaluator type. En-AU is the exception: annotators show no preference for SFTd over the baseline (48%, ties excluded) and weakly favour broad alignment over variety-targeted in T2 (34%). This divergence is consistent with Australian English features being more subtle due to its historical context (Collins, 2012) and, consequently, less perceptible to annotators, and reinforces that explicit adaptation does not uniformly improve perceived dialectal quality across varieties. Reward optimisation does not improve over supervised fine-tuning by any evaluator. When GRPO, the method achieving the highest automatic feature density, is compared directly against SFTd , neither Phi-4 condition nor human annotators show
Qualitative Analysis
Benchmark results alone do not reveal whether models produce dialectally marked output. We examine this through automatic metrics and human and LLM preference judgements, with representative qualitative examples shown in Table 2. 5.3
Human Evaluation and LLM-as-a-Judge
Reward Optimisation vs. Perceived Dialect Quality
As a training diagnostic, dialect feature density (Appendix E) shows that GRPO most aggressively optimises the reward signal, achieving the highest eWAVE feature density across all three families. Yet variety classifier accuracy (Appendix F), an independent signal covering all three model families, favours DPO for en-IN, with SFTd achieving the overall highest en-AU classifier accuracy. This divergence indicates that reward optimisation and holistic dialectal quality are not equivalent: the method that maximises the training objective does not necessarily produce outputs that are more recognisably dialectal by an independent measure. An independent, reward-free linguistic analysis cor-
6
Model
Prompt
Broad (-all)
Explicit (-d)
Qwen3-8B en-UK GRPO
What’s the best piece of advice someone’s ever given you?
“Be kind to others, and treat them the . . . This advice has helped us navigate way you would like to be treated.”. . . a through difficult situations and make defundamental principle of many cul- cisions that align with me core values. tures.
Llama-3.18B en-AU DPO
What are some small things that can brighten up a rough day?
. . . small things, depending on what . . . what makes you feel like makes you feel happy and ener- happy. . . Kindness to youself: Treat gized. . . Treat yourself to a small in- youself to a small indulgence. . . dulgence. . .
Gemma-34B en-IN GSPO
What’s your take on people who are always late?
. . . I understand that being late can be a frustrating experience for others. . . it’s not always appropriate to judge someone solely based on their punctuality. . .
. . . I am understanding that being late can be frustrating, especially where the punctuality is highly valued. . . it’s not appropriate to judge or are criticizing them without knowing the full contexts. . .
Table 2: Qualitative comparison of broad (-all) and explicit (-d) model responses. Broad models produce dialectally neutral outputs; explicit models exhibit surface dialectal features (bold). en-UK: possessive me, firstperson singular us. en-AU: like as discourse particle, youself. en-IN: stative progressive (am understanding), article overextension (the punctuality), progressive overextension (are criticizing). Examples are selected to illustrate clear surface dialectal features and are not claimed to be representative of all outputs at each pipeline stage.
a consistent preference in either direction across varieties. The high tie rate on Task 3 for en-IN (28 of 50 judgements excluded from the win rate) reflects the difficulty of distinguishing GRPOd from SFTd outputs for this variety, consistent with the finding that reward optimisation does not reliably produce outputs that are more recognisably dialectal. This is the clearest expression of the reward-quality gap: alignment pushes the reward signal upward, but that movement does not translate into outputs that evaluators find more dialectally convincing than the SFT baseline alone. Linking these judgements to output features, human preference tracks dialectal marking relative to the standard baseline but not reward density; annotators favour the more dialectally marked output over the instruct baseline yet prefer the more natural, contracted output over GRPO (Appendix J, Table 12). A direct comparison of the alignment methods shows that GRPO is the least preferred across all three varieties and both Phi-4 conditions, despite achieving the highest feature density. Human annotators prefer DPO for en-UK and en-IN, with Phi-4 in broad agreement; preferences for en-AU favour GSPO under both judge conditions. Full inter-annotator agreement statistics in Appendix I.
6
the performance ceiling on standard and dialectal benchmarks; alignment operates largely within that ceiling, producing small and inconsistent benchmark shifts while visibly changing generation, a pattern confirmed generation-side by an independent linguistic analysis: surface dialectal marking is established at supervised fine-tuning and only redistributed by alignment, remaining near-absent through the base, CPT, and instruct checkpoints (Appendix J, Figure 3). The clearest positive finding is that explicit variety targeting consistently produces output that human annotators and LLM judges recognise as dialectally distinct, and that variety-targeted alignment is preferred over broad alignment regardless of evaluator type (Task 2). En-AU is the exception, with annotators showing no clear preference at either stage. The reward-quality gap complicates this picture. GRPO achieves the highest eWAVE feature density across all three model families; yet on Llama-3.18B, where generation evaluation is conducted, it is the least preferred method in direct comparison across all three varieties and both Phi-4 conditions (Task 4). When compared directly against SFTd (Task 3), neither Phi-4 condition nor human annotators show a consistent preference for GRPO in either direction, meaning reward optimisation does not add evaluator-perceived value over the supervised baseline. DPO, which consistently falls below SFTd in feature density, achieves better variety classifier accuracy for en-IN and is preferred
Discussion
Our results highlight that dialectal robustness and generation are shaped by different pipeline components, and are not reliably co-indexed by any single method. CPT and SFT together determine 7
Human
Phi-4 (no ft.)
Phi-4 (+eWAVE)
T1
en-AU 48% (8t) en-IN 84% (7t) en-UK 87% (3t)
43% 71% 32%
81% 96% 60%
T2
en-AU 34% (9t) en-IN 71% (8t) en-UK 85% (4t)
100% 100% 100%
100% 100% 100%
T3
en-AU 46% (24t) en-IN 55% (28t) en-UK 38% (18t)
33% 32% 56%
44% 24% 40%
T4
en-AU GS (36%) GS (40%) GS (48%) en-IN D (50%) GS (36%) D (56%) en-UK D (52%) D (44%) D (52%)
Task Dial.
not to establish a single winning recipe, but to provide a controlled empirical account of what each pipeline component contributes and where current alignment methods succeed and fall short. Two findings stand out. The first is a robustnessgeneration dissociation: benchmark performance is shaped primarily by CPT and SFT, while alignment shifts generation in ways that benchmarks do not capture. The second is a reward-quality gap: on Llama-3.1-8B, the method that most aggressively optimises the dialectal reward signal is not the one preferred by LLM judges or human annotators; automatic evaluation across all three model families is consistent with this direction, though human and LLM preference evidence is limited to Llama-3.1-8B. These two findings together suggest that measuring dialect adaptation through benchmarks alone gives an incomplete and potentially misleading picture of what alignment achieves. The strongest positive result is that explicit variety targeting consistently produces output that human annotators and LLM judges recognise as dialectally distinct, and that variety-targeted alignment is consistently preferred over broad alignment regardless of evaluator type. This holds consistently for en-IN and en-UK; en-AU is the exception, with annotators showing no clear preference at either stage. No single alignment method dominates across conditions; results vary by model family, variety, and evaluation dimension, and the patterns we observe should not be treated as general prescriptions. Gemma shows greater sensitivity to CPT disruption than Llama or Qwen, and en-IN proves consistently more challenging than en-AU or en-UK across all methods. We release all model checkpoints and dialectal preference datasets to support future work. The varieties studied here remain substantially underresourced at the alignment stage, and the design of effective training signals for authentic dialectal generation is an open problem. DiaLLM provides a useful empirical foundation for understanding the current limits of dialect adaptation and a starting point for the work these varieties still need.
Table 3: Generation preference results across four tasks, three dialects, and three judge types. T1: Instruct vs SFTd (% SFTd wins, tie-excluded). T2: GRPOall vs GRPOd (% GRPOd wins, tie-excluded). T3: SFTd vs GRPOd (% GRPOd wins, tie-excluded). T4: preferred method and vote share (forced choice); D=DPO, GS=GSPO, GR=GRPO. Human: pooled across 2 annotators per dialect (50 judgements per task). Parenthesised t = tied judgements excluded from win rates. Bold: ≥60%.
by human annotators for en-UK and en-IN. This inversion points to a specific limitation of the reward basis: eWAVE is a typological atlas designed to document attested features across all 77 surveyed varieties, not to measure perceptual dialectal quality. Some features that receive high classifier weight overlap substantially across varieties or carry stronger associations with non-target varieties in general language use, so optimising the feature density objective can increase surface feature counts without producing output that is more recognisably dialectal to a human or LLM judge. Indian English proves consistently more challenging across most evaluation dimensions, likely reflecting the greater structural distance between IndE and the standard English training distribution and a denser active feature set (38 features) that creates more surface area for reward optimisation without producing recognisable output.
7
Conclusion
Limitations and Ethical Considerations
We presented DiaLLM, a dialect adaptation pipeline that continually pretrains three openweight LLM families on 18 English varieties from the International Corpus of English, then applies two post-training paradigms with a systematic comparison of DPO, GRPO, and GSPO. Our goal was
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-
8
specific improvement. BESSTIE covers our target varieties but is limited to sentiment and sarcasm, leaving many dimensions of dialectal competence unassessed. Second, eWAVE was designed as a typological database for classifying variety inventories, not as a reward signal for generative training. Some attested features are pragmatically conditioned, tending to surface only in specific conversational registers or construction types that may not be elicited by generic prompts. Others represent population-level tendencies that individual speakers vary substantially on, making binary detection inherently noisy. Operationalising these features as a learnable reward therefore introduces a mismatch: the signal favours outputs that superficially satisfy classifier-detectable patterns over those that are authentically dialectal in ways the inventory was never designed to measure. Beyond this, our reward function and generation evaluation share the same eWAVE-derived feature inventory, so feature density gains—retained in Appendix E as a training diagnostic only—reflect alignment with the training feature space rather than independently verified authentic dialect use. The BESSTIE-trained variety classifier provides results subject to register biases in the underlying corpus, as evidenced by the en-UK classifier’s near-total collapse to en-AU predictions on LLM-generated outputs (Section 5.3). The ICE corpus consists primarily of formal and semi-formal texts—legal proceedings, academic writing, scripted broadcasts—and our models may not generalise to informal or social media registers. Our pipeline is also limited to English dialects; extension to other languages would require new CPT corpora, feature inventories, and evaluation resources. Generation evaluation is conducted on Llama-3.1-8B only; whether the qualitative findings and preference results generalise to Qwen or Gemma remains to be established. Human preference evaluation used two annotators per variety (six total). Annotators were not dialect specialists and were selected on the basis of self-reported familiarity with the target variety. All judgements were conducted blind to model identity. Overall inter-annotator agreement is modest, reflecting the inherent difficulty of perceptual dialectal judgements with a small annotator pool; full statistics are reported in Appendix I. En-AU agreement is weakest (mean AC1 0.02), consistent with the weak preference results for that variety, and en-AU human preference results should therefore be treated as indicative rather than conclusive.
Finally, the implicit and explicit adaptation threads differ not only in whether dialectal perturbation is applied during SFT but also in the training data itself, with the implicit thread using UltraFeedback and the explicit thread using the dialectspecific preference dataset. Observed generation differences between threads therefore cannot be attributed solely to variety targeting, and future work isolating these factors through controlled ablation would sharpen the interpretation. Ethical Considerations This work aims to reduce linguistic inequities by improving LLM performance on under-represented English varieties. However, explicit dialect modelling carries risks: Multi-VALUE transformations may introduce stereotypical or overgeneralised patterns, and dialect-sensitive models could be misused to infer demographic or regional identity from text. DiaLLM should be interpreted as a step toward dialectal inclusivity rather than a complete mitigation of dialect bias. All model checkpoints and datasets released with this work will include documentation cards describing intended use, known limitations, and the dialectal communities represented.
References Marah Abdin, Jyoti Aneja, Harkirat Behl, Sébastien Bubeck, Ronen Eldan, Suriya Gunasekar, Michael Harrison, Russell J. Hewett, Mojan Javaheripi, Piero Kauffmann, James R. Lee, Yin Tat Lee, Yuanzhi Li, Weishung Liu, Caio C. T. Mendes, Anh Nguyen, Eric Price, Gustavo de Rosa, Olli Saarikivi, and 8 others. 2024. Phi-4 technical report. Preprint, arXiv:2412.08905. Emily M. Bender, Timnit Gebru, Angelina McMillanMajor, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’21, page 610–623, New York, NY, USA. Association for Computing Machinery. Su Lin Blodgett, Lisa Green, and Brendan O’Connor. 2016. Demographic dialectal variation in social media: A case study of African-American English. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1119–1130, Austin, Texas. Association for Computational Linguistics. Peter Collins. 2012. Australian english: Its evolution and current state. International Journal of Language, Translation and Intercultural Communication, 1:75– 86.
9
Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Wei Zhu, Yuan Ni, Guotong Xie, Zhiyuan Liu, and Maosong Sun. 2023. Ultrafeedback: Boosting language models with high-quality feedback. Preprint, arXiv:2310.01377.
Valentin Hofmann, Pratyusha Ria Kalluri, Dan Jurafsky, and Sharese King. 2024. Dialect prejudice predicts ai decisions about people’s character, employability, and criminality. Preprint, arXiv:2403.00742. Quzhe Huang, Mingxu Tao, Chen Zhang, Zhenwei An, Cong Jiang, Zhibin Chen, Zirui Wu, and Yansong Feng. 2023. Lawyer llama technical report. Preprint, arXiv:2305.15062.
Karin De Langis, Ryan Koo, and Dongyeop Kang. 2024. Dynamic multi-reward weighting for multi-style controllable generation. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6783–6800, Miami, Florida, USA. Association for Computational Linguistics.
Aditya Joshi, Raj Dabre, Diptesh Kanojia, Zhuang Li, Haolan Zhan, Gholamreza Haffari, and Doris Dippold. 2025. Natural language processing for dialects of a language: A survey. ACM Computing Surveys, 57(6):1–37.
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics.
Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. The state and fate of linguistic diversity and inclusion in the NLP world. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6282–6293, Online. Association for Computational Linguistics.
Fahim Faisal, Orevaoghene Ahia, Aarohi Srivastava, Kabir Ahuja, David Chiang, Yulia Tsvetkov, and Antonios Anastasopoulos. 2024. Dialectbench: A nlp benchmark for dialects, varieties, and closelyrelated languages. Preprint, arXiv:2403.11009.
Eeham Khan, Firas Saidani, Owen Van Esbroeck, Richard Khoury, and Leila Kosseim. 2025. Lowresource dialect adaptation of large language models: A French dialect case-study. Preprint, arXiv:2510.22747.
Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac’h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, and 5 others. 2023. A framework for few-shot language model evaluation.
Bernd Kortmann and Kerstin Lunkenheimer, editors. 2013. The electronic world atlas of varieties of English (eWAVE). Max Planck Institute for Evolutionary Anthropology, Leipzig. Bernd Kortmann, Kerstin Lunkenheimer, and Katharina Ehret, editors. 2020. eWAVE. Klaus Krippendorff. 2004. Content Analysis: An Introduction to Its Methodology, 2nd edition. Sage, Thousand Oaks, CA.
Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad AlDahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.
Joel Mire, Zubin Trivadi Aysola, Daniel Chechelnitsky, Nicholas Deas, Chrysoula Zerva, and Maarten Sap. 2025. Rejected dialects: Biases against African American language in reward models. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 7468–7487, Albuquerque, New Mexico. Association for Computational Linguistics.
Sidney Greenbaum and Gerald Nelson. 1996. The international corpus of english (ice) project. World Englishes, 15(1):3–15.
Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Preprint, arXiv:2305.18290.
Suchin Gururangan, Ana Marasović, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. Don’t stop pretraining: Adapt language models to domains and tasks. Preprint, arXiv:2004.10964.
Ricardo Rei, José G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and André F. T. Martins. 2022. COMET-22: Unbabel-IST 2022 submission for the metrics shared task. In Proceedings of the Seventh Conference on Machine Translation, pages 578–585.
Kilem L. Gwet. 2008. Computing inter-rater reliability and its variance in the presence of high agreement. British Journal of Mathematical and Statistical Psychology, 61(1):29–48. William Held, Caleb Ziems, and Diyi Yang. 2023. TADA: Task agnostic dialect adapters for English. In Findings of the Association for Computational Linguistics: ACL 2023, pages 813–824, Toronto, Canada. Association for Computational Linguistics.
Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. Preprint, arXiv:1908.10084.
10
David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. 2023. Gpqa: A graduate-level google-proof q&a benchmark. Preprint, arXiv:2311.12022.
353–355, Brussels, Belgium. Association for Computational Linguistics. Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, Tianle Li, Max Ku, Kai Wang, Alex Zhuang, Rongqi Fan, Xiang Yue, and Wenhu Chen. 2024. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. Preprint, arXiv:2406.01574.
Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. The risk of racial bias in hate speech detection. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1668–1678, Florence, Italy. Association for Computational Linguistics.
Tinakon Wongpakaran, Nahathai Wongpakaran, Derek Wedding, and Kilem L. Gwet. 2013. A comparison of Cohen’s Kappa and Gwet’s AC1 reliability coefficients for the same concordance data. Journal of Research in Nursing, 18(6):562–570.
Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint, arXiv:2402.03300.
Zedian Xiao, William Held, Yanchen Liu, and Diyi Yang. 2023. Task-agnostic low-rank adapters for unseen English dialects. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7857–7870, Singapore. Association for Computational Linguistics.
Dipankar Srirag, Aditya Joshi, and Jacob Eisenstein. 2025a. Predicting the target word of game-playing conversations using a low-rank dialect adapter for decoder models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers), pages 8–17, Albuquerque, New Mexico. Association for Computational Linguistics.
Qianqian Xie, Qingyu Chen, Aokun Chen, Cheng Peng, Yan Hu, Fongci Lin, Xueqing Peng, Jimin Huang, Jeffrey Zhang, Vipina Keloth, Xinyu Zhou, Lingfei Qian, Huan He, Dennis Shung, Lucila OhnoMachado, Yonghui Wu, Wang Qi, and Jiang Bian. 2024. Me-llama: Medical foundation large language models for comprehensive text analysis and beyond. Research square.
Dipankar Srirag, Aditya Joshi, Jordan Painter, and Diptesh Kanojia. 2025b. Besstie: A benchmark for sentiment and sarcasm classification for varieties of english. Preprint, arXiv:2412.04726.
An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388.
Mirac Suzgun, Nathan Scales, Nathanael Schärli, Sebastian Gehrmann, Yi Tay, Hyung Won Chung, Aakanksha Chowdhery, Quoc V. Le, Ed H. Chi, Denny Zhou, and Jason Wei. 2022. Challenging big-bench tasks and whether chain-of-thought can solve them. Preprint, arXiv:2210.09261.
Jiawei Zhao, Zhenyu Zhang, Beidi Chen, Zhangyang Wang, Anima Anandkumar, and Yuandong Tian. 2024. GaLore: Memory-efficient LLM training by gradient low-rank projection. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 61121–61143. PMLR.
Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penchev, and 197 others. 2025. Gemma 3 technical report. Preprint, arXiv:2503.19786.
Chujie Zheng, Shixuan Liu, Mingze Li, Xiong-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, Jingren Zhou, and Junyang Lin. 2025. Group sequence policy optimization. Preprint, arXiv:2507.18071.
Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. Llama: Open and efficient foundation language models. Preprint, arXiv:2302.13971.
Caleb Ziems, Jiaao Chen, Camille Harris, Jessica Anderson, and Diyi Yang. 2022. Value: Understanding dialect disparity in nlu. Preprint, arXiv:2204.03031. Caleb Ziems, William Held, Jingfeng Yang, Jwala Dhamala, Rahul Gupta, and Diyi Yang. 2023. MultiVALUE: A framework for cross-dialectal English NLP. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 744–768, Toronto, Canada. Association for Computational Linguistics.
Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural language understanding. In Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages
11
A
Corpus Details
ICE Variety
basis of both the dialect density metric and the masked reward used in dialect-specific adaptation. For reward masking, we use attested feature subsets for each target variety: 20 features for AusE, 38 for IndE, and 30 for NorthE. Table 5 presents a representative subset of features used for dialectspecific reward masking across the three target varieties: Northern British English (NorthE), Indian English (IndE), and Australian English (AusE). For dialect-specific reinforcement learning, reward computation is restricted to feature subsets corresponding to the target variety, based on attested feature–dialect mappings.
Cleaned Tokens Retention (%)
Australia Canada Great Britain Ireland USA India Singapore Malaysia Nigeria Philippines Cameroon East Africa Gibraltar Hong Kong Jamaica Scotland (spoken only) Trinidad & Tobago Uganda
655,336 1,439,016 1,002,845 1,342,878 637,093 1,570,211 1,412,390 382,839 1,308,130 1,486,351 393,254 1,105,083 539,240 1,899,811 1,396,602 584,070 1,403,912 1,355,122
100.0 38.0 100.0 61.7 63.5 41.5 49.1 96.7 100.0 45.3 74.0 99.8 78.3 37.1 40.6 100.0 60.6 34.3
Total (18 varieties)
19,916,246
52.4
C
All CPT and post-training experiments were conducted on NVIDIA A100 GPUs using a SLURMmanaged cluster, with the accelerate library for distributed execution. All runs use bfloat16 precision, gradient checkpointing, and a fixed random seed of 1234. Continual pretraining uses GaLore (Zhao et al., 2024) with AdamW over 64,000 steps, with a GaLore rank of 1024, scale 0.25, and a subspace update interval of 500 steps. The learning rate is 1 × 10−4 with linear warmup over 10% of training, weight decay of 0, gradient clipping at 1.0, and an effective batch size of 512. For both SFT and SFTd , models are trained for 3 epochs on chosen completions only in a single-turn instruction format, with a learning rate of 2 × 10−5 , cosine scheduling, and 5% warmup. All DPO runs use 5,000 steps with the sigmoid loss objective and β = 0.1. Gemma uses AdamW at 5 × 10−6 ; Llama and Qwen use 8-bit AdamW at 2 × 10−6 . Maximum prompt and completion lengths are 2048/256 for Gemma and Llama, and 1024/192 for Qwen, with a 5% warmup ratio. GRPO and GSPO share the same configuration, differing only in their importance sampling level: GSPO applies it at the sequence level, GRPO at the token level. Both use 4 generations per prompt over 5,000 steps, with sampling temperature 0.7 and top-p = 0.95. We set β = 0.02, ϵ = 3 × 10−4 , and ϵhigh = 4 × 10−4 . Reward weights are (0.8, 0.1, 0.1) for the dialect, COMET, and cosine components respectively; all components are z-score normalised using an exponential moving average with decay 0.99 and clipping at ±5. Optimiser settings and sequence lengths follow the DPO configuration, with weight decay 0
Table 4: ICE varieties used for continual pretraining, with retained token counts after preprocessing. Raw token total: 38,012,253. Retention varies due to differences in formatting, transcription quality, and completeness.
We source data from 18 varieties in the International Corpus of English (ICE), selecting those with sufficiently clean and accessible text (i.e., minimal markup corruption and complete transcript availability). Where required, access was obtained by contacting individual ICE teams, and only varieties for which usable data was available were included. The corpus spans a diverse set of regional English varieties, including both ’Inner Circle’ (e.g., Great Britain, USA) and ’Outer Circle’ (e.g., India, Nigeria, Singapore) contexts. Preprocessing removes markup, incomplete transcripts, and duplicated content. Retention rates vary substantially across corpora, reflecting differences in data quality and preprocessing requirements. For Scotland, only spoken data is included due to inconsistencies in the written portion. The final combined corpus contains approximately 20 million tokens. This corpus forms the sole source of data for continual pretraining across all models.
B
Hyperparameters and Infrastructure
Dialect Feature Inventory
Our dialect feature classifier is based on features derived from the eWAVE typological database, spanning 135 morphosyntactic and lexical phenomena across English varieties. These features form the 12
Dialect
Feature
Area
Rtg
Example
NorthE
Object pronouns as possessives (1SG) Us for singular object referent Was/were generalisation Nowt/owt (nothing/anything)
Pronouns Pronouns Agreement Adverbs
A A B B
He’s me brother Show us them boots he were thirsty I’ve got nowt
IndE
Progressive with stative verbs Definite article for indefinite Would for future No subject–aux inversion (wh-)
Tense/aspect Noun phrase Modality Complementation
A A A A
I’m liking this I had the toothache I would eat tomorrow What you are doing?
AusE
Progressive with stative verbs Bare adverbs (adj as adv) Discourse marker (sentence-final) Like as quotative/focuser
Tense/aspect Adverbs Discourse Discourse
B B B B
I’m liking this drive slow it was good, like she was like. . .
Table 5: Representative eWAVE-derived features used in the dialect feature classifier and reward function. Ratings follow the eWAVE attestation scale: A (pervasive), B (common), C (rare), D (limited structural context); X = not applicable, ? = no data.
and gradient clipping at 1.0. We selected λ by running full training for Llama 3.1-8B across all implicit and explicit threads with λ ∈ {0.50, 0.66, 0.80} and inspecting the composite reward trajectories. λ = 0.50 and λ = 0.66 showed faster saturation of the meaningpreservation components relative to the dialectal reward, while λ = 0.80 produced the most stable joint trajectory across threads. We therefore fixed λ = 0.80 for all model families and alignment methods.
D
features: sharpening underconfident predictions (e.g., T = 0.76) and softening overconfident ones (e.g., T = 1.78). This calibration ensures that sigmoid probabilities used as reward signals are well-scaled and comparable across features.
E Dialect Feature Density Across Pipeline Stages (Training Diagnostic) Table 6 reports dialect feature density as a training diagnostic: it characterises how each pipeline stage and alignment method optimises the reward signal, but is not used as a primary evaluation metric (see Section 4.3). Results cover all three model families, averaged across the 25 evaluation prompts.
Dialect Feature Classifier Details
The classifier is a BERT-base encoder (Devlin et al., 2019) with a lightweight linear projection head, trained using binary cross-entropy loss. Training data comprises approximately 148,000 sentences derived from standard NLU benchmarks (Wang et al., 2018) and transformed into dialectal variants using Multi-VALUE (Ziems et al., 2023); feature labels are obtained deterministically from the transformation process. On a held-out test set, the classifier achieves moderate overall performance, with reliable detection of high-frequency morphosyntactic markers and weaker coverage of rare or contextdependent features. Prior to reward computation we apply perfeature temperature calibration to the sigmoid outputs, performing an independent grid search over T ∈ [0.5, 3.0] for each of the 135 features on a held-out validation set to minimise Expected Calibration Error (ECE). The resulting temperature distribution is centred around unity (µ = 0.994, σ = 0.108), with targeted corrections for specific
Family
CPT
SFT
DPO
GRPO
GSPO
Broad thread (-all) Gemma 2.474 2.817 Llama 2.297 2.890 Qwen 2.834 2.932
2.912 2.939 2.950
2.959 2.950 2.951
2.949 2.911 2.941
Explicit thread (avg. across en-AU, en-IN, en-UK) Gemma — 1.570 1.504 1.628 1.617 Llama — 1.537 1.469 1.598 1.552 Qwen — 1.553 1.495 1.590 1.561
P Table 6: log(1 + σ(logits)) over active eWAVE features across pipeline stages (training diagnostic only). Broad thread uses all 135 features; explicit thread uses variety-specific subsets. Bold indicates highest alignment method per row. CPT is not applicable to the explicit thread as it precedes dialect-specific branching.
Three observations are consistent across all three families. First, every pipeline stage contributes dialectal signal: the log-sum score increases monotonically from CPT through alignment in the broad 13
thread, confirming that each component adds measurable reward-side signal. Second, the explicit thread is lower in absolute terms than the broad thread at equivalent stages, reflecting the smaller active feature subsets; the relative ordering across alignment methods is preserved. Third, GRPO achieves the highest score in the explicit thread across all three families, while DPO consistently falls below SFTd , a pattern discussed in Section 5.3.
F
Variety
Model
SFTd
DPO
GRPO
GSPO
en-AU
Llama Qwen Gemma
88.0 88.0 92.0
88.0 76.0 84.0
84.0 80.0 80.0
76.0 76.0 80.0
en-IN
Llama Qwen Gemma
48.0 52.0 56.0
64.0 76.0 60.0
48.0 36.0 60.0
40.0 36.0 56.0
en-UK
Llama Qwen Gemma
0.0 0.0 8.0
4.0 0.0 0.0
0.0 0.0 0.0
0.0 4.0 0.0
Table 7: Variety classifier accuracy (%) for the explicit thread over 25 evaluation prompts. Bold indicates best result per model per variety. en-UK accuracy is near zero across all families and methods, reflecting classifier failure rather than genuine dialectal signal (see Section 5.3).
Variety Classifier Accuracy Across Model Families
Table 7 reports variety classifier accuracy (%) for the explicit thread across all three model families, covering the four post-SFT stages. Accuracy is the proportion of the 25 evaluation prompt responses classified as the target variety. en-UK accuracy is near zero throughout all families and methods, confirming the classifier failure noted in Section 5.3 and Limitations; en-UK generation quality is therefore assessed via human and LLM-as-judge evaluation only. For en-AU and en-IN, DPO tends to maintain or improve classifier accuracy relative to SFTd , while GRPO and GSPO are generally lower—consistent with the finding that higher feature density does not correspond to stronger variety classifier signal. The en-UK variety classifier fails to generalise to LLM-generated outputs, with over 80% of responses classified as en-AU regardless of alignment method. This reflects a register distribution shift rather than a training failure: the classifier was trained on sentiment and sarcasm data rather than open-ended conversational responses. Accordingly, en-UK generation quality is not assessed via classifier accuracy and relies on human and LLM judgements instead.
G
Implicit Thread Benchmark Results
Qwen is largely robust to CPT and gains on several dialect-sensitive metrics, while Gemma is most affected and fails to fully recover DID after SFT. Alignment effects are model- and benchmarkspecific rather than systematic: Qwen BBH under DPO not only recovers from CPT and SFT degradation but exceeds the base model, suggesting that dialectal alignment can in some conditions enhance general reasoning, though the basis for this gain is not established and it is not replicated across families. Conversely, Llama BBH regresses notably under GRPO and GSPO, and Gemma DID collapses under DPO. Llama DID under GSPO is a notable exception where alignment adds meaningfully over SFT on a dialect task.
H
Human Annotation
We recruited two native human annotators for each variety, with the expectation that each would spend less than 2 hours on the full set of tasks. They consented to the use of their annotation data for this research. Any annotators who were not authors of this work have been paid for their time. All annotators were male, except one female annotator for en-UK. H.1
Task Instructions
We provided detailed instructions to our annotators on screen. The annotations process was divided into 4 tasks with 25 prompts per task. Overview: You will complete 100 judgements across 4 tasks (25 each). Each task
14
Qwen3-8B
Llama 3.1-8B
Gemma 3-4B-it
Benchmark Base CPT SFT DPO GRPO GSPO Base CPT SFT DPO GRPO GSPO Base CPT SFT DPO GRPO GSPO BBH GPQA GLUE VALUE BST-Sent BST-Sarc DID
70.0 55.9 53.7 74.15 36.6 37.2 36.2 38.4 68.4 59.1 64.0 63.6 56.1 50.3 65.6 64.8 86.1 89.9 90.2 90.1 37.5 42.8 41.5 41.3 56.4 61.3 64.9 65.8
73.0 36.4 63.9 66.2 90.2 40.5 64.6
73.0 39.9 64.3 66.0 90.2 41.3 65.8
46.5 45.4 45.6 48.0 31.4 31.0 27.8 29.3 52.9 46.6 49.3 49.5 43.2 38.8 43.8 44.1 87.1 87.3 84.7 85.4 38.3 34.9 38.4 39.0 38.9 40.6 59.3 50.4
30.7 28.8 49.2 43.7 85.3 38.5 57.1
34.1 30.3 48.9 42.8 85.1 38.9 61.3
48.7 41.0 40.3 45.4 27.8 30.3 29.5 33.8 62.8 50.6 50.8 51.0 52.8 42.3 43.7 45.1 85.2 82.2 82.5 78.5 34.4 12.5 25.8 34.2 59.4 35.1 26.4 13.7
45.4 29.8 52.6 45.5 84.4 24.5 34.6
45.5 30.8 51.1 43.4 82.4 26.0 27.7
Table 8: Implicit adaptation thread results (%). Base, CPT, and SFT are sequential pipeline stages; DPO, GRPO, and GSPO are alignment methods applied on top of the SFT checkpoint without dialect-specific targeting. Bold indicates the best result within each model family per row. BST-Sent and BST-Sarc refer to BESSTIE sentiment and sarcasm respectively.
asks you to read responses to a prompt and judge which sounds more dialectal for your assigned variety of English. Full instructions appear at the start of each task.
• What do you make of people who are always late? • What are some good ways to wind down at the end of a long day? • What’s a piece of advice that tends to really stick with people?
For Tasks 1 to 3, instruction snapshot is quoted below.
Opinion • Do you think it’s better to live in the city or the countryside? • What do you think about people who just order takeaway every night instead of cooking? • Is it worth spending a lot of money on a gym membership, or can you stay fit without one? • What’s the most overrated tourist destination in your opinion? • Do you think social media does more harm than good?
You will see 25 pairs of responses to the same prompt. Select whichever response sounds more dialectal for en-X English — closer to how a speaker of that variety might naturally respond. Here X is replaced by language variety. For Task 4, however, these were changed to: You will see 25 sets of three responses to the same prompt. Select whichever response sounds most dialectal for en-X English. There is no tie option - pick the best of the three. Here X is replaced by language variety. H.2
Food & Lifestyle • What’s a simple meal worth recommending to someone just learning to cook? • What’s a good thing to eat on a lazy Sunday morning? • Any tips for eating well without spending a fortune? • What’s a good breakfast to start the day right? • What’s a dish that everyone should try at least once?
Generation Prompts
The following 25 prompts are used across all generation evaluation methods (automatic metrics, human preference evaluation, and LLM-asjudge). Prompts are manually curated to be openended, conversationally natural, and free of dialectanchored geography, technical content, or sensitive topics.
Hypothetical • If you could only eat one meal for the rest of your life, what would it be and why? • If you had a free weekend with no plans or obligations, how would you spend it? • If you could live anywhere in the world, where would you pick and why? • If you could only keep three apps on your phone, which would you choose?
Casual chat • What’s a good comfort meal after a rough day? • What do you reckon makes a good night out with mates?
15
• If a friend asked you to help them move house, what would you bring to make the day easier?
Variety
Task
AC1
Kripp. α
Cohen’s κ
Obs.%
Personal Reflection • What’s a hobby worth picking up, and what makes it enjoyable? • What are some small things that can brighten up a rough day? • What would an ideal Saturday look like from morning to night? • What’s a TV show or film worth recommending to everyone? • What’s a surprisingly useful thing to know that most people don’t think about?
en-AU
T1 T2 T3 T4
−0.02 −0.03 0.25 −0.11
0.00 −0.18 0.34 −0.06
−0.02 −0.17 0.20 −0.06
32 28 48 24
en-IN
T1 T2 T3 T4
−0.04 0.17 0.54 −0.25
−0.14 0.08 0.45 0.01
−0.15 0.00 0.21 0.00
28 40 64 12
en-UK
T1 T2 T3 T4
0.42 0.66 0.57 0.11
0.36 0.43 −0.19 0.24
0.36 0.27 −0.08 0.16
60 72 64 40
I
Table 9: Inter-annotator agreement per variety and task. T1: Instruct vs SFTd ; T2: GRPOall vs GRPOd ; T3: SFTd vs GRPOd ; T4: DPO/GSPO/GRPO forced choice. en-UK T3 illustrates the prevalence paradox: AC1 0.57 despite α −0.19 and κ −0.08, owing to near-unanimous rater preference for SFTd .
Inter-Annotator Agreement
Inter-annotator agreement (Table 9) is modest overall, reflecting the inherent difficulty of perceptual dialectal judgements with a small annotator pool. Agreement is highest on Task 3 (AC1 0.25–0.57; observed 48–64%), where raters consistently agree that the dialect-adapted output is more marked, consistent with the perceptibility finding in Section 5.4. Agreement is near chance on Tasks 1 and 4 (AC1 ≤ 0.42, mostly near zero), reflecting that the differences alignment introduces are small and hard to distinguish, consistent with the reward–quality gap. en-UK shows the strongest overall agreement (mean AC1 0.44), including strong agreement on T2 (AC1 0.66, 72% observed), the one task where raters showed stronger consensus; enAU agreement is weakest (mean AC1 0.02), consistent with the weak preference results for that variety. Because several tasks are dominated by a single outcome, standard chance-corrected statistics are deflated by the prevalence paradox; we therefore report Gwet’s AC1 as the primary statistic alongside Krippendorff’s α and Cohen’s κ for transparency (Krippendorff, 2004; Gwet, 2008; Wongpakaran et al., 2013).
J
evaluation outputs (three families × full pipeline × 25 prompts × variants). Detectors. We detect three independent signal types: (i) lexical markers from curated per-variety lexicons (e.g. en-AU arvo, en-IN prepone, enUK nowt); (ii) orthographic markers via explicit British/American spelling pairs (a classifierindependent en-UK/en-AU signal); and (iii) morphosyntactic features over spaCy POS/dependency parses – progressive-with-stative (I am understanding), was/were generalisation, possessive me, pluralised mass nouns (informations), bare adverbs, and invariant tags. Detectors are high-precision by design; reported rates are therefore lower bounds. From the firings we derive density (features per 1k tokens), diversity (distinct feature types), and stacking (features per sentence); the latter two are not captured by the density-based reward. The reward–quality gap is corroborated by an independent measure, across families. Table 11 and Figure 2 show that, although GRPO maximises the eWAVE reward density (Table 6), it produces the fewest independent surface markers for Llama and Qwen (e.g. Qwen density 0.87 vs. SFTd 2.34 per 1k tokens) and the lowest feature diversity; the ordering is mixed only for Gemma. These results suggest the eWAVE reward does not straightforwardly correspond to independently measured surface dialectal richness, extending the reward– quality gap (Sec. 5.3) beyond Llama to all three families on a non-circular measure.
Dialectal Generation: Linguistic Analysis
The reward signal and the variety classifier (Sec. 5.3) both derive from the same eWAVE feature space, so neither is an independent check on whether adapted models actually produce dialectal language. We therefore analyse the generation outputs with a separate, rule-based instrument that does not reuse the reward classifier. All measures are computed by a released script over the 1,375
16
1.50 1.25 1.00 0.75 0.50 0.25 0.00
eWAVE reward density (Table 6; higher = more reward) GRPO
SFTd
DPO
GRPO
GSPO
dialectal features /1k tokens
log-sum eWAVE density
Alignment redistributes, rather than introduces, dialectal marking. Figure 3 traces en-IN marker density along the pipeline: it is near zero through the base, CPT and instruct checkpoints, rises sharply at dialectal SFT, and is thereafter only shifted by the alignment methods, with GRPO (the method that most optimises the reward) sitting lowest. These results suggest surface dialectal marking is largely established by continual pretraining and supervised fine-tuning and subsequently redistributed by alignment, consistent with the robustness–generation dissociation (Sec. 5.3).
3.0
Independent surface markers (ours; /1k tokens, 95% CI)
2.5 2.0 1.5
GRPO
1.0 0.5 0.0
SFTd
DPO
GRPO
GSPO
Figure 2: eWAVE reward density (left; reproduced from Table 6) versus independent surface-marker density (right; ours, bootstrap 95% CIs), averaged over the explicit thread. GRPO is highest on the reward it optimises yet lowest on independent markers—the reward–quality gap made visible.
Human preference tracks dialect, not reward density. Linking the measures to the human pairwise judgements on Llama (Table 12), annotators prefer the more dialectal output over the standard baseline (Task 1, ∆density +1.38, 95% CI excludes 0) but prefer the more natural, contracted output over the reward-maximising GRPO output (Task 2, ∆contractions +2.87; ∆density negative, CI includes 0). Preference for dialectal density therefore appears non-monotonic: features appear to help against standard English, but their rewarddriven maximisation does not, consistent with the gap reported in Sec. 5.3.
en-IN density /1k tokens
Dialectal marking across the pipeline (en-IN) standard
7 6 5 4 3 2 1 0 base
CPT
instruct
SFTd
DPO
GRPO
GSPO
Figure 3: Dialectal-marker density across the pipeline for en-IN. Marking is near zero through base, CPT and instruct (standard region), rises sharply at dialectal SFT, and is thereafter only shifted by alignment—GRPO, the reward-maximising method, lowest (bootstrap 95% CIs).
Surface marking is concentrated in en-IN. Table 10 shows that 37.7% of en-IN explicit-thread outputs carry a detected marker, against 10.0% for en-UK and 3.3% for en-AU; en-IN also shifts furthest from the standard pole. The varieties also differ in which features surface (Figure 4): enIN marking is carried by mass-noun pluralisation and progressive-with-stative, whereas the few enUK markers are predominantly possessive me and was/were generalisation. The en-AU markers we detect (bare adverbs, invariant tags) are essentially absent, consistent with the observation (Sec. 5.3) that Australian features are register-flexible and not reliably elicited by casual prompts.
Variety
% marked
Density (/1k)
JS vs. std.
en-AU en-IN en-UK
3.3 37.7 10.0
0.16 4.17 0.73
0.113 0.361 0.142
Table 10: Surface dialectal marking is concentrated in en-IN. % marked = share of explicit-thread outputs with ≥ 1 detected marker; JS = Jensen–Shannon divergence of the feature distribution from the standard pole (base/instruct), averaged over methods/families. en-AU markers (bare adverbs, invariant tags) are essentially absent, matching the register-flexibility note in Sec. 5.3.
Caveats. Detectors favour precision over recall, so densities are lower bounds; en-AU/en-UK are sparse and quantitative density claims are reliable mainly for en-IN; the human bridge is Llama-only with modest n (48–85 non-tie pairs per task), so deltas are reported with bootstrap 95% CIs as directional evidence. Code, detectors, and lexicons are released for reproduction.
17
total feature firings
Feature-type composition by variety (explicit thread)
200 175 150 125 100 75 50 25 0
mass-noun plural prog.+stative British spelling
en-AU
possessive me lexical was/were
en-IN
en-UK
Figure 4: Feature-type composition by variety (total firings, explicit thread). en-IN marking is carried by mass-noun pluralisation and progressive-with-stative; en-UK by possessive me and was/were generalisation; en-AU markers are essentially absent.
Family
Method
Density (/1k)
Diversity (types)
Stacking (/sent)
Llama
SFTd DPO GRPO GSPO
1.51 1.77 1.24 1.50
0.19 0.19 0.15 0.19
0.028 0.033 0.023 0.027
Qwen
SFTd DPO GRPO GSPO
2.34 2.32 0.87 1.40
0.25 0.28 0.13 0.19
0.043 0.043 0.014 0.026
Gemma
SFTd DPO GRPO GSPO
2.36 1.79 1.69 1.44
0.21 0.25 0.20 0.19
0.042 0.033 0.031 0.028
Table 11: Independent surface-marker measures on the explicit thread, averaged over en-AU/en-IN/en-UK (means over 25 prompts per cell). Density = curated dialectal features per 1k tokens; Diversity = distinct feature types; Stacking = features per sentence. In contrast to the eWAVE reward density (Table 6, where GRPO is highest), GRPO yields the lowest independent density and diversity for Llama and Qwen – the linguistic signature of the reward–quality gap (Section. 5.3).
Task T1: instruct vs SFTd T2: broad vs targeted T3: SFTd vs GRPOd
∆Density ∆Stack ∆Contr. +1.38 +0.78 -0.51
+0.03 +0.01 -0.01
n
-1.16 85 -0.02 83 +2.87 48
Table 12: Human-preference bridge (Llama). Mean winner−loser delta in each linguistic measure across non-tie pairwise trials; bold = bootstrap 95% CI excludes zero. Positive = humans prefer the output with more of the measure. T4 is excluded as a forced threeway choice is not amenable to winner−loser delta computation. Humans prefer the more dialectal output over standard (T1) but the more natural, contracted output over reward-maximising GRPO (T3): preference for dialectal density is non-monotonic, consistent with the reward–quality gap (Section. 5.3).
18