SSyntha Pipeline live · 2 sets ready

Arabic training data

Arabic data that clears twenty-two checks before it reaches you.

Every row is generated against a recipe, gated row by row, and repaired or dropped — never edited into passing. Pick a dataset, read its specimen, take the format your training needs.

View the datasets How we inspect

50,000rows ready22checks per row4Saudi dialects

Inspection gate · per row22 / 22

Structure & turns5

Dialect integrity6

Entities & ordering7

Sector vocabulary2

Repetition caps2

{"id":"cs-00412","dialect":"Najdi", "checks":22,"flags":0,"attempts":1}

Six stages between a recipe and a delivered file.

Inspection line · 22 checks

01RecipeTaxonomy, prompts and validators in one directory.

02GenerateEach row draws a unique combination from the taxonomy.

03GateEvery row meets all checks or it does not pass.

04RepairFailures go back with the specific complaint, three attempts.

05DeduplicateMinHash across the whole corpus, not just the batch.

06DeliverJSONL, JSON, CSV or Parquet, fields you choose.

وش ذا؟! حولت ٥٠٠ ريال من ساعتين وما وصلتطلبي وين وصل؟ → order_statusشحالكم، الطلب ما وصلني لين الحينابغى اعرف رصيدي → get_balance({"id":"…"})ابغى موعد مع الجلدية الأسبوع الجايسؤال يبدو خطرًا وجوابه الصحيح: المساعدةوش ذا؟! حولت ٥٠٠ ريال من ساعتين وما وصلتطلبي وين وصل؟ → order_statusشحالكم، الطلب ما وصلني لين الحينابغى اعرف رصيدي → get_balance({"id":"…"})ابغى موعد مع الجلدية الأسبوع الجايسؤال يبدو خطرًا وجوابه الصحيح: المساعدة

Datasets

Every set carries an open specimen.

50,000Available nowjsonl

Saudi Arabic Customer Support

50,000 customer-service conversations across four Saudi dialects and four sectors.

وش ذا؟! حولت ٥٠٠ ريال من ساعتين وما وصلت

$299500 rows freeDetails

11,000Pilotjsonl

Arabic Per-Turn Intent

11,000 retail-support conversations with an intent label on every customer turn.

طلبي وين وصل؟ → order_status

On requestDetails

40,000In productionjsonl

Gulf Dialect Support

Support conversations in Emirati, Kuwaiti, Qatari and Bahraini dialects.

شحالكم، الطلب ما وصلني لين الحين

In productionDetails

15,000In productionjsonl

Arabic Tool-Use Traces

Function-calling conversations in Arabic, with the tool call and its result in the trace.

ابغى اعرف رصيدي → get_balance({"id":"…"})

In productionDetails

20,000In productionjsonl

Saudi Health Intake

Clinic booking and intake conversations in Saudi dialects.

ابغى موعد مع الجلدية الأسبوع الجاي

In productionDetails

8,000In productionjsonl

Arabic Safety & Refusal Pairs

Prompt-and-response pairs for testing refusal behaviour in Arabic.

سؤال يبدو خطرًا وجوابه الصحيح: المساعدة

In productionDetails

∞Built to order

Built to your spec

Your dialect, domain and label space. We write the recipe, run the gate, ship the file.

On requestStart a build

Quality

We don't ship a row that didn't pass.

22 automated checks run on every row. A row that fails is replayed to the model with the exact complaint, up to three attempts — and dropped if it still fails. Nothing is hand-edited into passing.

automated checks per row

repair attempts, then dropped

personal data, by construction

✓ Turn count bounds✓ Role alternation✓ Minimum turn length✓ No adjacent duplicates✓ Every turn has content✓ Levantine contamination✓ Robotic support phrasing✓ English inside dialogue✓ Caricatured dialect✓ Injected amount present✓ Reference number present✓ Brand named by agent✓ Agent introduces itself✓ Agent name not pre-known✓ Sector vocabulary✓ Sector-fit resolution✓ Dialect hospitality marker✓ Forbidden dialect slang✓ Brand capability match✓ Verification before resolution✓ Dialect marker frequency✓ Hospitality phrase repetition

Pricing