G0m7s/deliverius-ai

Deliverius synthetic dataset (1.0.0)

Synthetic data for Deliverius, a fictional delivery app, and Deli, its customer-support agent: customer messages labelled by intent, a refund-claim history labelled for fraud, policy questions with the section that answers them, and end-to-end scenarios with the expected final state. Every label is true by construction: it comes from the scenario the code drew, never from a language model.

Synthetic data notice

Everything here is made up. Customers, restaurants, orders and messages describe no real person, company or event. Names, CPF numbers (they have valid check digits but belong to nobody), phone numbers, e-mail addresses and street addresses are generated with Faker. They exist to test personal-data masking and must not be treated as real data. The prompt-injection sentences inside some messages are test data.

How it was generated

Five layers, in order. Only the fourth uses a language model.

1. A simulator draws customers (honest, opportunist or fraudster profiles), restaurants and six months of orders with delivery timelines and incidents.
2. Claims come from the profile and the incidents; the fraud label says whether the claim has no real incident behind it. A small fraction of labels is flipped on purpose.
3. The expected resolution of each claim is computed by the same policy engine the application uses.
4. A language model writes the wording of customer messages from the facts of a drawn scenario. It sees only markers such as [[CPF]]; the code fills them with fake values and records where each personal datum landed. The intent label is the scenario's.
5. End-to-end scenarios are built without a model, with the expected final state taken from the policy engine.

• Generator version: 1.0.0, seed 42, commit 2ea250c5421de8e8fd14a30682cee45c0b829030.
• Reference instant of the simulated period: 2026-10-31T23:59:59+00:00.
• Faker version: 40.44.0.

Role	Model
datagen	gemini:gemini-3.5-flash-lite
datagen_holdout	groq:openai/gpt-oss-120b

Prompt	Version	SHA-256
holdout	v1	9d98cba49237d49464a25c2ad3a2e33edd8eba1392a796ae4a5f801023fb8a98
messages	v1	7fff4523ae0c1a7f9c16d15835a2590c1b1d45ce5550d9bf28aeeaf04c50b2ba
rag_pairs	v1	3e373a338d6354f21b33f3403b0e6f48f1f7d34e727d34584e3cfff1247d0d17

Sizes and splits

Set	Rows
customers	2000
e2e_scenarios	150
e2e_scenarios.F1	30
e2e_scenarios.F2	25
e2e_scenarios.F3	20
e2e_scenarios.F4	25
e2e_scenarios.F5	20
e2e_scenarios.F6	15
e2e_scenarios.F7	15
failed_scenarios.messages	0
failed_scenarios.messages_holdout	0
failed_sections.policy_qa	0
fraud_ground_truth	3378
fraud_history.test	498
fraud_history.train	2880
messages	2969
messages.test	440
messages.train	2089
messages.validation	440
messages_holdout	550
orders	20000
policy_qa	60
restaurants	60
scenarios.messages	297
scenarios.messages_holdout	55

Messages are split 70/15/15 by scenario, stratified by intent: the variations of one scenario are always on the same side.

The fraud history is split by time: the test set is everything at or after 2026-10-01T00:00:00+00:00. The fraud rate among claims is 11.8% and the fraction of flipped labels is 2.5%.

Messages per intent:

Intent	messages	messages.test	messages.train	messages.validation	messages_holdout
item_missing	270	40	190	40	50
wrong_item	270	40	190	40	50
quality_issue	270	40	190	40	50
late_delivery	270	40	190	40	50
not_delivered	270	40	190	40	50
billing_issue	270	40	190	40	50
cancel_order	269	40	189	40	50
coupon_question	270	40	190	40	50
policy_question	270	40	190	40	50
sensitive	270	40	190	40	50
other	270	40	190	40	50

Fields

• messages: text, intent (one of 11), split, scenario_id, pii_spans (kind, start, end of each personal datum in text), has_injection, persona.
• messages_holdout: The same fields as messages, written by a model of another family from scenarios that share nothing with the main set. Test only.
• fraud_history: One claim per row: order, customer, instant, reason, claimed items and amount, the policy resolution, and is_fraud. No customer profile and no flipped-label mark: those are in fraud_ground_truth/, for analysis only.
• policy_qa: question, and the document and section of the policy that answers it.
• e2e_scenarios: flow (F1 to F7), the initial_state (customer, order, prior actions), the customer's goal, and the expected final state.

Known limitations

• The data is synthetic: rates, prices and behaviours are plausible, not calibrated against a real service.
• The intent of a message is the scenario's. The model may write something vaguer, or off topic; that is also what real messages do, and the manual review below counts it.
• All messages of one set come from one model, so their style is less varied than that of real customers. messages_holdout exists to measure how much a classifier memorised that style.
• Portuguese (Brazil) only.
• Manual review defect rate: 5.0% (sample of 100 messages).
• Discard rate of messages variations: 3.0% overall (91 discarded, 2969 kept); highest for not_delivered: 10.0%.
• Discard rate of messages_holdout variations: 0.0% overall (0 discarded, 550 kept); highest for billing_issue: 0.0%.

License

Creative Commons Attribution 4.0 International (CC BY 4.0). The generator version, seed and commit above reproduce the files.