mballouch/jev-darija-telecom

Jev-Darija Telecom: Moroccan Darija Customer-Service Messages

19,457 synthetic customer messages in Moroccan Darija (Arabic script, Arabizi, and Darija mixed with French), written as customers of Moroccan mobile and internet operators would write them, and labelled with a two-level telecom intent taxonomy: 12 categories, 75 intents, plus product, urgency and a needs-human flag.

No public Darija telecom intent dataset existed, so this one was generated with large language models: mostly Claude Opus 5.5 (89.8%), with Gemini 3.8 Flash (10.2%) for extra data on the hardest intents. It is synthetic and has not yet been reviewed by native speakers.

Splits

Split	Messages	Arabic script	Arabizi	Mixed (Darija + French)
train	18,350	9558 (52%)	4574 (25%)	4218 (23%)
test	1,107	592 (53%)	293 (26%)	222 (20%)

Training has 188–350 messages per intent (more for intents a first model found hard); test has 13–15 per intent.

How it was generated

Generator	Messages	What it wrote
Claude Opus 5.5	17,475 (89.8%)	Round 1: 150 training + 15 test messages per intent. Round 2: 50–100 more training messages per intent, about half of them written to separate intents a first model confused. All test messages.
Gemini 3.8 Flash (via OpenRouter)	1,982 (10.2%)	Round 3: about 100 more training messages for each of 23 of the weakest intents.

The generator and round fields record this for every message.

All writers followed the same written guide: everyday Darija rather than Modern Standard Arabic or translation; customer vocabulary (الريزو، الصولد، نشارجي، لافيبر، الروتور…); varied spelling, openings, tones (calm to angry), lengths (2-word messages to multi-sentence stories) and writer personas from different Moroccan cities; concrete details (dirham amounts, durations, what the customer already tried); no real names, phone numbers (written 06XXXXXXXX) or account numbers. Each intent's messages were written with its confusable neighbours in mind.

Automatic checks applied to every message:

• valid labels, and a script label that matches the text (Arabizi has no Arabic letters, Arabic-script text is not mostly Latin, mixed text contains French; no Persian/Urdu letters);
• near-duplicates removed within each intent (character-trigram Jaccard ≥ 0.8), including cross-script restatements (an Arabizi retelling of an Arabic message, compared on consonant skeletons);
• no training message close to a test message of the same intent, in the same or another script.

The test split was written separately: different writer personas and cities (Oujda, Tétouan, Laâyoune, El Jadida, Errachidia, Safi) that never appear in training, without reference examples. It was frozen after round 1, and later rounds dropped any training message that overlapped it.

Taxonomy

Full definitions, examples, and the product, urgency and needs-human criteria are in telecom_taxonomy.json. 26 intents correspond to the intents of the English Bitext telco dataset (taxonomy only; no Bitext text is included); the rest cover the Moroccan market: prepaid recharges and credit, data passes, PUK codes, scam SMS, fibre eligibility, line registration, the fixed home line, TV, mobile money wallets, entertainment add-ons and loyalty points.

Category	Intents	Names
billing_payment	8	invoice_info, invoice_dispute, payment_methods, pay_bill, schedule_payments, payment_not_reflected, refund_compensation, paid_services_charges
prepaid_credit	5	recharge_how, recharge_failed, balance_check, balance_deducted, credit_transfer_loan
usage_mobile_data	7	check_usage, excess_charges, usage_limits, data_pass_buy, data_not_working, data_slow, hotspot_apn
home_internet	7	install_internet, fiber_eligibility, installation_delay, home_outage, home_slow, router_wifi, tv_issue
home_line	7	relocation, home_line_activation, home_line_suspend_restore, ownership_transfer, adsl_to_fiber_migration, fixed_phone_issue, equipment_return
network_calls	8	check_signal_coverage, report_poor_signal, call_quality_drop, cannot_call_receive, sms_issue, international_calls, activate_call_features, deactivate_call_features
sim_line	8	activate_line, deactivate_line, sim_blocked_puk, lost_stolen, sim_replacement_esim, line_identification, change_provider, scam_spam_report
plans_subscription	9	plan_offers_info, sign_up_plan, change_plan, cancel_plan, cancellation_fee, promotions_bonus, roaming, entertainment_services, loyalty_points
devices	3	device_purchase, device_repair_warranty, device_settings
customer_care	7	file_complaint, complaint_followup, human_agent, customer_service_contact, agency_location, account_app_access, greeting_thanks
mobile_money	5	mm_send_receive, mm_cash_in_out, mm_payment, mm_transaction_issue, mm_account_access
out_of_scope	1	non_telecom

Fields

Field	Description
id	Message id
text	The customer message
script	arabic, arabizi, or mixed (Darija with French)
category	One of 12 categories
intent	One of 75 intents (always inside category)
product	prepaid_mobile, postpaid_mobile, mobile_broadband, home_adsl, home_fiber, home_box_4g_5g, fixed_phone, tv, mobile_money, digital_services, or unknown when the message does not say
urgency	0 information · 1 problem, service still works · 2 service down or money at stake · 3 service down and angry, or an emergency
needs_human	Whether the message should go to a human agent (dispute, fraud, theft, explicit request, repeated unresolved problem)
persona	The fictional writer the message was written as (e.g. "taxi driver, Casablanca")
generator	claude-opus-5.5 or gemini-3.8-flash
round	Generation round (1, 2, or 3)

License

CC BY 4.0.