Output formats
We deliver in your format
Default delivery is JSONL — one record per line, Hugging Face-ready. We also produce Label Studio JSON, SRT/VTT for audio, and custom CSV or TSV on request.
JSONL
One annotation record per line. Compatible with Hugging Face datasets, OpenAI fine-tuning and most ML pipelines out of the box.
{"id": "ax-001",
"text": "שלום, אני צריך עזרה",
"lang": "he",
"dialect": "modern",
"label": "support_request",
"confidence": 0.97,
"annotator": "native_he_001",
"reviewed": true}
RLHF Preference Pairs
Structured for DPO / PPO training. Each record contains a prompt, chosen response and rejected response with annotator rationale.
{"id": "ax-rlhf-044",
"prompt": "اشرح البرمجة للمبتدئين",
"chosen": "البرمجة هي ...",
"rejected": "برمجة = كود ...",
"lang": "ar", "dialect": "msa",
"rationale": "chosen: clearer, formal",
"iaa_score": 0.88}
Audio / Speech
WAV + JSONL metadata. Audio segmented by utterance, with speaker tag, dialect label, start/end timestamps and transcription.
{"id": "ax-audio-017",
"file": "ax-017.wav",
"start_ms": 0, "end_ms": 3200,
"speaker": "F_AR_EG_02",
"dialect": "eg",
"text": "عايز أعرف إمتى ...",
"code_switch": false,
"snr_db": 42.1}
HTR / OCR
Image + ground-truth text pairs. PNG/JPEG image of the handwritten or printed region, with UTF-8 transcription and bounding-box coordinates.
{"id": "ax-htr-209",
"image": "page_04_line_03.png",
"text": "ירושלים עיר הקודש",
"script": "hebrew_print",
"bbox": [12, 44, 320, 72],
"confidence": 0.94,
"era": "modern"}
Language coverage
Seven dialect tiers — one vendor
Most vendors cover Modern Standard Arabic. We cover the dialects your users actually speak — and Hebrew, which almost no Arabic-focused vendor touches.
| Language / Dialect | Code | Speaker pool | Tasks available | Notes |
|---|---|---|---|---|
| Modern Hebrew | he-IL |
✓ Active | All | Primary market — including code-switch with English |
| Egyptian Arabic | ar-EG |
✓ Active | All | Largest dialect speaker base in our pool |
| Levantine Arabic | ar-LEV |
✓ Active | Annotation · Audio · LLM eval | Syria · Lebanon · Jordan · Palestine |
| Gulf Arabic | ar-XG |
✓ Active | Annotation · Audio | Saudi Arabia · UAE · Kuwait · Bahrain |
| Modern Standard Arabic | ar-MSA |
✓ Active | All | Used for formal, news and educational content |
| Moroccan Arabic (Darija) | ar-MA |
⬤ Limited | Annotation · Audio | Requires scoping call — longer lead time |
| Hebrew–Arabic Code-Switch | he+ar |
⬤ Specialist | Audio · LLM eval | Rare combination; dedicated pool of bilingual speakers |
Quality & inter-annotator agreement
We measure quality before you do
Every delivery includes a QA report. Batches that do not reach our IAA threshold are re-annotated internally — you never receive a batch below the agreed bar.
Minimum inter-annotator agreement for binary and multi-class tasks. We use Cohen's Kappa for two annotators comparing a single categorical label.
Used for ordinal, interval and multi-annotator tasks — including RLHF rating scales. Tolerates missing data and scales to more than two annotators.
At least 10% of every batch is re-annotated by a senior reviewer before delivery. For pilots, this rises to 25%.
First-pass acceptance rate across active projects. Measured as batches delivered without a client-requested re-do.
QA workflow — every batch
- Annotator 1
Primary annotation against task guidelines
- Annotator 2
Independent second pass — no visibility of A1
- IAA check
Automated disagreement flagging. Items below κ = 0.6 go to adjudication
- Adjudication
Senior linguist resolves conflicts and documents the decision rationale
- QA sample
10–25% blind re-annotation by a separate QA reviewer
- Delivery + report
JSONL batch + QA metrics PDF sent together
Task taxonomy
What we annotate
NER
Named Entity Recognition
Person, organisation, location, date, product — in Hebrew and all Arabic dialects. Supports nested and overlapping spans.
SENT
Sentiment & Tone
Positive / Negative / Neutral at sentence or document level. Extended schemes include intensity (1–5) and specific emotions.
RLHF
RLHF Preference Pairs
Two model responses ranked by a native speaker. Includes a rationale field. Used for DPO, PPO and Constitutional AI pipelines.
CLASS
Intent & Topic Classification
Single or multi-label. Taxonomies provided by you or built jointly. Calibration set included in every pilot.
AUDIO
Speech Transcription
Verbatim or clean-read. Speaker diarisation, dialect labelling, code-switch flagging and confidence scores included.
HTR
Handwriting Transcription
Hebrew historical and modern scripts. Bounding boxes at line or word level. Ground-truth text in NFD-normalised UTF-8.
SAFE
Safety & Red-Teaming
Harmful content detection, cultural sensitivity checks and adversarial prompt evaluation in Hebrew and Arabic.
AUDIO-COLLECT
Speech Data Collection
Scripted and spontaneous recordings. Controlled dialect, age, gender distribution. WAV 16kHz 16-bit, noise screened.
Delivery & SLAs
Timelines you can plan around
Scoping call
Response to first contact. We schedule a 30-min call within two business days.
Pilot proposal
After the scoping call — fixed price, scope, IAA target and acceptance criteria.
Pilot launch
From signed agreement to first annotated items. Includes annotator briefing and calibration set.
Pilot delivery
Typical 100–500 item pilot. Delivered with full QA report and IAA breakdown.
Volume batches
Once at volume — batches delivered weekly with rolling QA metrics. Daily for urgent projects.
Re-annotation SLA
If a delivered batch fails your QA, we re-annotate flagged items within five business days at no charge.
Sample dataset
See the data before you commit
Request a sample of up to 50 annotated items — covering the task type and language you need. Delivered in JSONL within 72 hours, under NDA if required.
- Choose language and dialect
- Choose task type (NER, RLHF, audio, etc.)
- Receive sample + IAA report within 72 hrs
- No commitment to a pilot or contract
NDA available on request