NajahUniv/arabic-univeristy-chatbot-qa

Arabic University Chatbot QA

A multilingual, multi-label intent-routing dataset for a university chatbot: given a student's message, predict which of 20 intent categories it should route to. This is routing, not question answering — the dataset contains no answers.

Release v0.8.0 — pinned as a Hub tag, so revision="v0.8.0" always resolves to exactly these rows.

This release holds 50,000 question rows in 22,600 scenario groups. Every row has accepted == true; the classifier target is verified_labels.

Languages

Language	language value	Rows	Share
Colloquial Arabic (dialect)	colloquial_arabic	22,207	44.4%
Modern Standard Arabic	msa_arabic	13,899	27.8%
English	english	6,441	12.9%
Arabizi (Arabic in Latin script)	arabizi	4,153	8.3%
Hebrew	hebrew	2,523	5.0%
Arabic-English code-switching	mixed	777	1.6%

Why Hebrew is included

Hebrew is included to reflect the university community this dataset serves. Students and staff include Arabic-speaking residents of cities such as Nazareth, Jaffa, and Acre (Akko), many of whom also speak Hebrew. Including Hebrew helps represent the languages users may naturally choose when seeking university information and services.

Dialects

arabic_dialect records the dialect a scenario group was written in. Shares below are of the 33,474 rows that name one.

Dialect	arabic_dialect value	Rows	Share
Palestinian or broadly Levantine (pre-control rows)	palestinian_or_levantine	11,080	33.1%
Egyptian	egyptian	5,770	17.2%
Palestinian	palestinian	4,172	12.5%
Gulf / Khaleeji (Saudi, Emirati, Kuwaiti, Qatari)	gulf	3,672	11.0%
Iraqi	iraqi	2,972	8.9%
Maghrebi (Moroccan, Algerian, Tunisian)	maghrebi	1,837	5.5%
Sudanese	sudanese	1,456	4.3%
Levantine (Jordanian, Syrian, Lebanese)	levantine	1,422	4.2%
Yemeni	yemeni	1,093	3.3%

A further 16,526 rows carry unspecified, meaning the row's own text carries no dialect at all: it is Modern Standard Arabic, code-switching, or not Arabic.

Labels

Each row carries one to four labels. out_of_scope never appears with another label.

Label	Covers
admissions	university applications, acceptance, requirements, documents, procedures, and application deadlines; exclude course registration after admission
academic_programs	majors, degrees, diplomas, specializations, available programs, and choosing a field; exclude individual courses and registration procedures
registration	course registration, add/drop, closed sections, prerequisites, withdrawals, and registration restrictions; exclude university applications
tuition_and_payments	tuition, credit-hour costs, balances, invoices, payment procedures, installments, refunds, payment methods, and financial holds
scholarships_and_aid	scholarships, grants, financial assistance, exemptions, discounts, eligibility, and aid applications; exclude ordinary payment questions
academic_calendar	semester dates, registration dates, holidays, exam periods, withdrawal deadlines, and semester start or end dates
exams_and_grades	exams, schedules, grades, transcripts, GPA, academic standing, objections, failed courses, and incomplete grades
courses_and_study_plans	course descriptions, study plans, curriculum requirements, prerequisites, credit hours, electives, compulsory courses, and course sequences
transfer_and_equivalency	transfer students, university or program transfers, course equivalency, recognition of prior study, and credit transfer
graduation	graduation requirements, applications, clearance, certificates, diplomas, commencement, and eligibility
student_services	student affairs, ID cards, official documents, clubs, activities, counseling, complaints, and general non-academic support
technical_support	login and password problems, account access, portal errors, university email, Moodle, broken links, and technical malfunctions
library	library hours, borrowing, book availability, databases, journals, digital resources, library accounts, and research-resource access
research_and_postgraduate	master's or doctoral research, theses, supervisors, research centers, ethics, publications, postgraduate procedures, and research support
staff_and_departments	faculty members, employees, colleges, departments, responsibilities, staff profiles, and the responsible administrative unit
contact_and_location	phone numbers, email addresses, office or campus locations, maps, directions, working hours, and ways to contact a person or unit
campus_services	transportation, housing, cafeterias, clinics, parking, accessibility, sports facilities, and other physical campus services
news_and_events	university news, announcements, conferences, workshops, ceremonies, lectures, activities, and upcoming events
general_university_information	university history, leadership, administration, rankings, accreditation, mission, statistics, campuses, and general institutional facts
out_of_scope	requests unrelated to university information or services; always alone

Variants

Each scenario group holds one original question and one or more variants that share its group_id and its meaning. variant_mode distinguishes them: original, reword (a rewording in the same language), and translation (the same request in another language or dialect). 1,896 paraphrases are faithful translations of their original into another language or dialect rather than rewordings. All variants of a group are always assigned to the same split, so a paraphrase can never leak across the train/test boundary.

Splits

Split	Groups	Rows
train	18,080	40,023
validation	2,260	4,990
test	2,260	4,987

Splits are assigned by deterministic multi-objective stratification over labels, language, meaningful language/dialect combinations, routing mode, difficulty, noise, and confidence. A language/dialect combination present in at least three independently assignable components is guaranteed to appear in every split.

Columns

Column	Meaning
question	The user message. This is the classifier input.
verified_labels	JSON array of intent labels. This is the classifier target.
generated_labels	Labels the generator intended, before independent verification.
uuid	Stable row identifier, group_id:variant_index.
group_id	Shared by every variant of one scenario; splits never break a group.
group_size	Number of variants in the group.
variant	original for the first question, paraphrase for the rest.
variant_index	Position within the group; 0 is the original.
variant_mode	original, reword, or translation; see Variants above.
language	Language of this row's own question.
source_language	Language of the group's original question.
question_length	Requested length band: terse, medium, detailed, or unspecified.
arabic_dialect	Dialect the group was written in, e.g. palestinian, egyptian, gulf. Dialect-bearing rows generated before the dialect became a per-run control carry palestinian_or_levantine; unspecified means the row's own text carries no dialect.
routing_plan	The intent plan the group was generated against.
routing_mode	single_label or multi_label.
target_category	Primary label of the routing plan.
noise_level	Requested typo/imperfection level.
difficulty	Requested classification difficulty.
paraphrase_style	Rewording transformation requested for the group's paraphrases.
confidence	Verifier confidence in this row's labels, 0-1.
equivalence_confidence	Verifier confidence that the group's variants match, 0-1.
semantic_equivalence	Whether the verifier judged every variant equivalent.
plan_conformant	Whether the group matched its routing plan.
semantic_issues	JSON array of issue codes; empty for accepted rows.
usable	Whether the verifier judged this row usable.
accepted	Always true in this release.
generator_model	Model that wrote the question.
verifier_model	Model that independently verified its labels.
generator_temperature	Sampling temperature used by the generator.
verifier_temperature	Sampling temperature used by the verifier.

How it was made

Every question is synthetic. A generator model wrote it, then a separate verifier model independently re-derived its labels without seeing the requested plan. Only groups where every variant was judged usable, equivalent, and plan-conformant with confidence >= 0.80 are included. Deterministic checks additionally reject questions whose text does not match its required language and rows whose generated and verified labels disagree.

Near-duplicate scenarios are collapsed to one representative using character 3-5 gram TF-IDF cosine similarity >= 0.90, so the published rows contain no known exact or near-duplicate cross-group pairs and none can leak across splits. A further 29 rows were dropped because near-identical questions carried conflicting labels.

Each row records the models and sampling temperatures that produced it, so rows added in later releases with different providers or models stay distinguishable:

• Generator: gpt-4.1-mini (34,483 rows), z-ai/glm-5.3-flash (10,409 rows), deepseek/deepseek-v4-pro-0813 (5,108 rows)
• Verifier: gpt-4.1-mini (50,000 rows)

Rows carried over from before those columns existed are recorded as gpt-4.1-mini at generator temperature 0.7 and verifier temperature 0.0, and carry question_length unspecified. Their dialect is not unknown: the generator asked every scenario for "Palestinian or broadly Levantine" before the dialect became a per-run control, so the dialect-bearing ones among them are recorded as palestinian_or_levantine.

Full provenance, removal counts, split fingerprints, and distributions are published in merge_report.json.