VohoAI/voho-saudi-dialogues

Voho Saudi Dialogues

13,156 multi-turn conversations in spoken Saudi Arabic (Najdi), 105,808 turns, 694,240 words, from Voho. Apache 2.0.

Two halves. 7,603 service calls across the eight sectors Voho's voice agents work in — a technician handing over a rig shift, a customer disputing a SADAD charge, a permit-to-work request — and 5,553 everyday conversations between people who know each other: family, food, driving, weddings, the Hilal–Nassr match. Nothing like the first half existed in the open before; the second half is what keeps a model trained on the first from treating every sentence as a support ticket.

This is the training data behind VohoAI/voho-saudi-chat-4b.

What is in it

Every dialogue cleared two independent checks before inclusion:

1. A Najdi lexicon gate — Saudi function words present (وش، تبي، الحين، ليش، زين) and their Modern Standard Arabic equivalents absent (ماذا، أريد، الآن، لماذا).
2. MARBERTv2, an Arabic dialect classifier — at least 60% of a dialogue's replies must read as Gulf. Across the kept set, 89% of replies do, against 94.5% for a held-out reference of Saudi text.

Nothing with markdown, tables or code, no reply longer than a spoken turn, no duplicated openings. Dialogues run 6–10 turns, mean reply 7 words — phone-call length.

Service calls (set = enterprise)

Sector	Dialogues
oil-gas	1,062
telecom	1,007
banking	950
utilities	935
logistics	928
healthcare	917
real-estate-facilities	916
government	888

Four scenarios per sector, four staff roles (agent, field technician, shift supervisor, receptionist), eight caller personas.

Everyday (set = everyday)

Topic	Dialogues
shopping	546
food	524
travel	491
occasions	478
work	475
friends	474
sports	472
family	434
home	426
driving	416
health	412
weather	405

Four scenarios per topic, eight relationships (friend, sibling, colleague, neighbour, cousin, parent), eight speaker personas.

Example

oil-gas · تسليم وردية في موقع حفر

عميل: أنا صاحب المعدات. الوردية السابقة رجعت الحفار مكسور وما وقعت على الاستلام موظف: أي حفار؟ ووين موقع التسليم؟ عميل: حفار صغير. موقع ٢. الحين الوردية الجديدة تقول مو مسؤولين موظف: لا، هم مسؤولين. أعطني صورة العطل وأنا أرفع تقرير فوري عميل: أرسلت الصورة الحين. أبي حل قبل لا يروحون موظف: تم، فتحت تذكرة. بيجيك مندوب خلال ساعة يوقع معك

Schema

One JSON object per line:

{"id": "voho-0", "set": "enterprise", "topic": "oil-gas", "scenario": "تسليم وردية في موقع حفر",
 "messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

messages alternates user / assistant, always starting with user and ending with assistant, so it drops straight into a chat template. In the service calls user is the customer and assistant the staff member; in the everyday set user is whoever speaks first.

Use

from datasets import load_dataset
ds = load_dataset("VohoAI/voho-saudi-dialogues", split="train")
enterprise = ds.filter(lambda r: r["set"] == "enterprise")

Licence

Apache 2.0. Use it commercially, fine-tune on it, redistribute it.

Limitations

• Najdi. Central Saudi. Hijazi and Khaleeji speakers will find it close but not theirs.
• Written, not transcribed. This is dialogue text in a spoken register, not transcripts of recorded calls. It has no disfluencies, overlaps or noise.
• The classifier's Gulf class covers the UAE and Kuwait too. A high Gulf score means "reads as Gulf", not "reads as Riyadh".
• Topic coverage is deliberate, not representative. Eight sectors and twelve everyday topics, evenly; it is not a sample of what Saudis talk about.
• Written without diacritics.