ZynoviqA sample of the data, and the schema it follows.
Five illustrative records on Indian GST, the fields every record carries, and the files that ship with a delivery. The records below are synthetic and shortened to a table; a delivered record carries the full reasoning and provenance.
Five illustrative records
Illustrative Synthetic records written to show the shape of the data. Not certified; not drawn from the benchmark.
| Id | Topic | Facts (summary) | Verified answer | Legal basis | Valid as of | Difficulty |
|---|---|---|---|---|---|---|
| GST-ITC-0001 v2 Illustrative | Blocked credits — proviso | Factory with 400 workers runs a canteen required by law; catering invoice carries ₹10,000 GST | Credit allowed: ₹10,000 | CGST s.17(5)(b) proviso; OSH Code, 2020 (canteen where 100 or more workers); Circular 172/04/2022 | FY 2026–27 | Hard |
| GST-ITC-0002 v1 Illustrative | Blocked credits — plain | Office with 40 employees provides catered lunch voluntarily; no legal obligation to provide it | Credit blocked | CGST s.17(5)(b)(i) | FY 2026–27 | Easy |
| GST-ITC-0003 v1 Illustrative | Conditions for credit — GSTR-2B | Registered buyer holds a valid tax invoice and has paid the supplier; the supplier has not reported the invoice and it does not appear in the buyer's GSTR-2B | Credit not available until the invoice is reported and appears in GSTR-2B | CGST s.16(2)(aa) | FY 2026–27 | Medium |
| GST-ITC-0004 v1 Illustrative | Blocked credits — passenger vehicles | Company hires passenger cars with a seating capacity of up to 13 for executive travel; not used for transporting passengers as a business, training or any purpose required by law | Credit blocked | CGST s.17(5)(b)(i) read with s.17(5)(a) | FY 2026–27 | Medium |
| GST-RCM-0005 v1 Illustrative | Reverse charge — legal services | A firm of advocates supplies legal services to a registered business entity in India | Recipient pays GST under reverse charge; the firm does not charge GST on the invoice | Notification 13/2017–Central Tax (Rate), as amended | FY 2026–27 | Medium |
The schema, field by field
Every record in our standard schema carries these fields. When you specify your own schema, the same content is mapped into it.
| Field | Type | What it holds |
|---|---|---|
id | string | Stable identifier for the case; never reused |
version | integer | Increments whenever the answer, reasoning or legal basis is revised |
domain · topic · subtopic | string | Position in the domain topic map, for example GST → input tax credit → blocked credits |
task_type | enum | Determination, computation, classification, posting, document review |
facts | text | The scenario as a professional would receive it; synthetic |
documents[] | array | Optional synthetic invoices, vendor records or similar, as text or file references |
question | text | What the AI is asked to determine |
answer | object | Structured fields: decision (yes/no or code), amount, section, with accepted alternatives where more than one answer is correct |
outcome_conditions[] | array | The conditions that change the outcome: which fact, if different, would change the answer, and to what |
reasoning[] | array | Ordered steps, each with the rule applied at that step |
legal_basis[] | array | Section, notification, circular or ruling, with citation text |
valid_as_of | date or period | The point in law the answer is correct for; triggers revision when the rule changes |
difficulty | enum | Easy, medium, hard, ambiguous |
common_mistake | text | The plausible wrong answer and why it is wrong |
author_qualification · reviewer_qualification | string | Professional qualification of each contributor, by anonymous id |
reviewer_agreement | enum | Unanimous, majority, adjudicated |
adjudication_status | enum | Not required, adjudicated, ambiguous (competing positions recorded) |
dissent | text | Present only on ambiguous records: the minority position and its authorities |
provenance | object | How the record was made and checked; timestamps; checks passed |
permitted_use | enum | Training, evaluation, internal, as set by your licence |
tags[] | array | Evaluation tags, for example reverse-charge, proviso, time-limit |
What ships with every delivery
| File | Contents |
|---|---|
| data files | The records, in JSONL, CSV or Parquet, in your schema or ours |
| provenance.json | Per record: version, valid-as-of date, anonymous contributor ids, reviewer agreement, adjudication status |
| quality_report.pdf | Counts by topic and difficulty, agreement rates, how certification was done |
| licence.json | Licence type, permitted uses, term, licensee, copy identifier |
| changelog.md | What changed since the previous version |
Delivered by secure download or direct transfer to your cloud storage, from Indian cloud regions. Never through code repositories. Formats, delivery and licences.
Downloads
JSONL and CSV sample files — Planned. Until they are published, we share a small sample in your schema on request before any full delivery.
How these records were made
A delivered record is written by a capability-qualified professional, solved independently by others, reviewed, adjudicated where they disagree, and tied to the written rule with a valid-as-of date. AI never sets, changes or certifies professional ground truth. How it works · The five kinds of data.
Sample dataset — FAQ
Can we download this sample as JSONL or CSV?+
Not yet. Downloadable sample files in JSONL and CSV are planned. Until then, we share a small sample in your schema before any full delivery, so your pipeline is tested early. Request a dataset.
Are these real records from the production system?+
No. They are illustrative records written to show the shape of the data. They use only facts already published on this site and are to be checked by a practising Chartered Accountant before publication as a certified sample.
Do delivered records look exactly like this?+
The fields do. The layout follows your schema when you specify one, and otherwise our documented standard schema. Formats, delivery and licences.
Is the sample drawn from the benchmark?+
No. Benchmark cases are held out and never shown as sample or sold as training data. Sample and training records come from the training pool. Training and test separation.