ZynoviqTrain your AI on how a qualified professional actually reasons.
General models learned mostly from material written elsewhere. On Indian tax, payroll and company law they sound fluent and are often wrong. We write the examples that close that gap: realistic tasks, verified answers and the reasoning behind them, each solved independently by qualified professionals and tied to the law.
Five kinds of professional data
| Data | What it contains | Typical use |
|---|---|---|
| Reasoning records | How a professional gets from the facts to the answer, step by step, with the rule at each step | Supervised fine-tuning |
| Preference data | Which of two answers a professional prefers, and why | Preference tuning and reward models |
| Edge and ambiguous cases | Provisos, exceptions and genuine disagreement, recorded with the dissent | Robustness; calibrating when to escalate |
| Failure data | Real failure patterns of AI on professional work, with the corrected answer | Targeted improvement |
| Evaluation sets | Held-out cases kept apart from all training data | Your own internal testing |
Every item is traceable
A certified record is not a question and an answer. It is a documented professional judgment you can audit.
| Field group | What the record carries |
|---|---|
| Identity | Stable id and version |
| Task | Facts and documents, as text or as synthetic invoices, vendor records and similar |
| Verified answer | Structured fields, with accepted alternatives where more than one answer is correct, and the conditions that would change the outcome |
| Reasoning | The steps a professional took, with the rule at each step |
| Legal basis | Section, notification or circular, with a valid-as-of date |
| Difficulty | Easy, medium, hard or ambiguous |
| People | Author and reviewer qualification, by anonymous id |
| Quality | Reviewer agreement and adjudication status |
| Provenance and use | How the record was made and checked; permitted use under your licence |
Two ways to buy
| Data Mission | Licensed packs and expert case subscription | |
|---|---|---|
| What it is | Built to order: topics, period, size, difficulty mix and schema set by you and approved in a written specification | Ready-made certified packs, or a recurring stream of newly certified cases |
| Rights | Licence; exclusive for a period; or ownership, by agreement | Non-exclusive |
| Kept current | Change notes for affected records when the law changes | Refreshed as the law changes |
| How it is priced | Fixed price against the approved specification, quoted | Annual licence or monthly subscription, quoted |
| Start | Request a dataset | Browse the catalogue |
Your training data will never be our test.
Training data and benchmarks come from the same production system but from separate, controlled pools. Held-out cases are never sold as training data, and training cases never enter a benchmark. A lab that buys our data can therefore be measured on a benchmark it has not seen. How the separation works.
How a Data Mission runs
- You describe the need
What your model does, where it is weak, the topics, the period of law, the size and the schema you want.
- Specification approved
A written specification covering topics, difficulty mix, evidence requirements, rights, delivery and acceptance criteria. Nothing starts until you approve it.
- Professionals mobilised and calibrated
Capability-qualified professionals in the specific topics, calibrated on a shared set before production begins.
- Written, solved, reviewed, adjudicated, tied to law
Each case written by one professional, solved independently by others, reviewed, adjudicated where they disagree, and tied to the written rule with a valid-as-of date.
- Delivered
Records in your schema, with provenance, quality and licence files. A typical first mission takes five to seven weeks.
Formats
JSONL, CSV or Parquet, in your schema or ours. Every delivery includes a provenance file, a quality report, a machine-readable licence and a change log. Formats, delivery and licences.
What we do not promise
We do not promise that our data improves every model. Improvement is measured per task, per model and per training method, and it depends on how you use the data. What we can do is measure it: run a held-out benchmark on the same topics before and after, and report what moved.
Buy the data with the benchmark that measures it.
For labs that want evidence, not a promise, we offer the two products together:
- A dataset on the topics you choose.
- A held-out benchmark on the same topics, from the separate pool, never in your training set.
- A baseline run on your model before delivery, and a re-run after you train.
Priced as a project fee plus a private benchmark run, quoted. By agreement, a small part of the fee can be tied to measured improvement on the held-out benchmark.
Offered to one lab at a time while the model is proven.
Who buys it
- AI labs and model builders — frontier labs improving performance on Indian professional work, and India's labs building models for India.
- Agent companies — teams building finance and compliance agents that need to reason, not just retrieve.
Professional Data — FAQ
What is a Data Mission?+
A dataset built to order. You describe what your model needs; we write a specification covering topics, period, size, difficulty mix, schema, rights and acceptance criteria; you approve it; qualified professionals write, solve, review and adjudicate the cases; we deliver the records with provenance, quality and licence files. Request a dataset.
Who owns the data?+
That is set in the specification. Most missions are licensed, exclusive for a period where you need it. Ownership can be transferred by agreement for data built only for you. Exclusivity is priced separately because it forecloses reuse. Formats, delivery and licences.
Can we see a sample first?+
Yes. The sample dataset shows illustrative records and the schema. Before any full delivery we also share a small sample in your schema, so your pipeline is tested early.
How do you stop the data leaking into tests?+
Training and held-out cases come from separate, controlled pools with separate access rights, duplicate detection across pools and answer keys in an isolated store. Your training data is never our test, and our test is never your training data. Training and test separation.
Is the data available in Indian languages?+
English today. Native Indian-language cases, certified by bilingual professionals, are planned from 2027, starting with Hindi.