Access is granted manually by the author. Please tell us who you are and how you intend to use the data; requests without a real name, email and intended use will be declined.

This dataset was generated with Claude Opus 4.6 (via Amazon Bedrock) and is released for research on Arabic dialect alignment. By requesting access you confirm that your use complies with the Anthropic Usage Policy and the Amazon Bedrock service terms, including the restriction on using model outputs to develop models that compete with Anthropic, and that you will not redistribute the data outside those terms.

Saudi Dialect DPO: Najdi & Hijazi preference pairs

22,170 preference pairs that teach a chat model to answer in the requested Saudi dialect (Najdi or Hijazi) without drifting into Modern Standard Arabic, the other Saudi region, or a non-Saudi dialect. Every pair is a controlled corruption: chosen and rejected carry the same content and length and differ only in dialect, so a DPO policy learns dialect fidelity rather than length, politeness, or verbosity.

Format is ready for TRL's DPOTrainer (conversational prompt / chosen / rejected).


Pairs	22,170 (11,117 Najdi · 11,053 Hijazi)
Words	5.2M Arabic words across prompts, chosen and rejected
Topics	18, each between 1,141 and 1,320 pairs
Task types	chat reply · customer service · constrained instruction · short QA · rewrite
Registers	peer (7,085) · elder (7,596) · customer (7,489)
Fault types	wrong Saudi region (10,724) · MSA drift (6,807) · foreign-dialect leak (4,639)
Mean length	user turn 19 words · chosen 105 · rejected 111
Generator / judge	Claude Opus 4.6 on Amazon Bedrock
Gate	LLM-as-judge rubric, numeric thresholds, one repair round

Why this dataset

Arabic LLMs default to MSA, and when asked for "Saudi" they usually produce a generic Gulf blend. Public Saudi resources are either country-level (no Najdi/Hijazi distinction), comprehension-only (multiple choice), or private. Existing fidelity metrics use a Gulf-vs-MSA classifier that cannot see Najdi words in a Hijazi reply. This dataset targets exactly that gap: sub-national dialect control with a preference signal that isolates the dialect variable.

What a pair looks like

{
  "prompt": [
    {"role": "system", "content": "تكلم باللهجة الحجازية السعودية، بصفتك موظف خدمة عملاء يرد على عميل."},
    {"role": "user",   "content": "ياخي حسابي انقفل فجأة بدون سبب وأنا محتاج أسحب فلوسي ضروري، ممكن تساعدني وش أسوي؟ رد علي بدون إنجليزي"}
  ],
  "chosen":   [{"role": "assistant", "content": "يا خوي لا تشيل هم، الموضوع له حل إن شاء الله. أول شي روح لأقرب فرع للبنك حقك ... وكلّم موظف خدمة العملاء هنيّ وقلّه إن حسابك انقفل فجأة وتبغى تعرف السبب ... لو ما قدرت تروح الفرع دحين ... لو تبغى أي شي ثاني قلّي يا قلبي."}],
  "rejected": [{"role": "assistant", "content": "ياخي لا تشيل هم ... وكلّم موظف خدمة العملاء هناك وقلّه إن حسابك انقفل فجأة وتبي تعرف وش السبب ... لو ما قدرت تروح الفرع الحين ... لو تبي أي شي ثاني تكفى قلّي يا بعد حيي."}],
  "meta": {"id": "p007502", "dialect": "hijazi", "topic": "finance", "task_type": "customer_service", "register": "customer", "fault": "wrong_region"}
}

The rejected reply keeps every fact and the same length; only the Hijazi markers (هنيّ، تبغى، دحين، يا قلبي) were swapped for Najdi ones (هناك، تبي، الحين، تكفى، يا بعد حيي).

Fields

field	type	description
`prompt`	list of messages	A short system tag (dialect + register) and the user turn, written in the target dialect as people type online.
`chosen`	one assistant message	Reply in the target dialect, obeying any constraints the user asked for.
`rejected`	one assistant message	Same reply re-voiced with one injected fault (see `meta.fault`).
`meta.id`	string	Stable id, shared with the generation logs.
`meta.dialect`	`najdi` \| `hijazi`	Target dialect of `chosen`.
`meta.topic`	string	One of 18 daily-life domains (government services, housing, health, food, …).
`meta.task_type`	string	`chat_reply`, `customer_service`, `constrained_instruction`, `short_qa`, `rewrite`.
`meta.register`	string	Who the user is addressing: `peer`, `elder`, `customer` (service agent).
`meta.fault`	string	`wrong_region` (other Saudi dialect), `msa_drift` (formal MSA), `foreign_leak` (Levantine / Egyptian / Iraqi markers).

The system tag is intentionally short (for example تكلم باللهجة النجدية السعودية، بصفتك صديق من نفس العمر.) so it can be reproduced at inference time. The long generation prompt used to produce the data is not part of the training record.

How it was built

Marker lexicon. 1,200 region-tagged marker words and phrases (Najdi, Hijazi, shared Saudi, MSA, foreign) drafted by Opus 4.6 and used to steer generation and to report marker density. It is not the quality gate.
Prompt bank. 23,558 user turns sampled uniformly over dialect × topic × task type × register, MinHash-deduplicated (442 near-duplicates removed). About a third carry machine-checkable constraints (sentence count, must-include phrase, no English, end with a question, numbered list) that the user asks for in natural words.
Chosen reply. Generated with a curated dialect feature guide and a forbidden list of cross-Gulf and non-Saudi words that LLMs tend to leak into "Saudi" text (وايد، مب، شلون، هلأ، بدي، إزيك، عايز …).
Rejected reply. A full re-voicing of the chosen reply with one fault, keeping meaning, sentence count, numbers and names. Fault mix: 50 % wrong region, 30 % MSA drift, 20 % foreign leak.
Judge gate. Opus 4.6 at temperature 0 scores each pair on a fixed rubric (chosen_dialect 1–5, chosen_natural 1–5, chosen_constraints_ok, rejected_fault_present, rejected_meaning_preserved 1–5, rejected_same_length). A pair survives only if every score is ≥ 4 / true. First-pass survival was 80 %.
Repair round. Failed pairs were regenerated once on the failing side with the judge's reason as feedback and re-judged; 3,339 of the final pairs came through repair. Pairs that failed twice were dropped (1,388 terminal drops).

Judge score distribution on the released pairs: chosen_dialect 4 → 20,390 · 5 → 1,780; chosen_natural 4 → 19,613 · 5 → 2,557; rejected_meaning_preserved 4 → 1,814 · 5 → 20,356.

Intended use

DPO / ORPO / KTO-style post-training of Arabic chat models for Saudi dialect control.
Evaluation of dialect drift: the rejected side is a labelled negative for each fault type.
The meta.fault field lets you train or evaluate on one failure mode at a time.

Limitations

Synthetic. Both sides of every pair are model-written. The dialect is Opus 4.6's rendering of Najdi and Hijazi, filtered by an Opus 4.6 judge. A model trained on it inherits that judge's taste; the judge was not calibrated against native raters at release time.
Two dialects only. Southern, Northern and Eastern Saudi varieties are absent.
User turns can mix. A small share of user turns tagged Hijazi contain Najdi words (and vice versa). The gate scores the assistant replies, not the user turn.
Marker-dense style. Chosen replies were steered toward explicit dialect markers, which biases style slightly toward marker-rich text.
Judge strictness. The judge drops replies it calls "leaning formal" even when a native speaker might accept them, so the released set is conservative.
No safety or factuality filtering was applied; replies are advisory everyday content.

Provenance and terms

All text was generated by Claude Opus 4.6 . Use is subject to the Anthropic Usage Policy and the Amazon Bedrock service terms, which restrict using model outputs to develop models that compete with Anthropic. Access is gated so that each user acknowledges this. The data license is CC BY-NC 4.0 for the compilation; it does not grant rights beyond those terms.

Citation

@dataset{haroon2026saudidpo,
  title   = {Saudi Dialect DPO: Najdi and Hijazi preference pairs by controlled corruption},
  author  = {Hesham Haroon},
  year    = {2026},
  publisher = {Hugging Face},
  url     = {https://huggingface.co/datasets/HeshamHaroon/saudi-dialect-dpo}
}

Related work by the same author

HeshamHaroon/saudi-dialect-conversations: 3,545 multi-turn Najdi SFT conversations.

Downloads last month: 22

Total file size:

59.6 MB